runtime: transduction is a language concern, so move it into the language

#141 let signal enter as geometry and it worked, but it was placed at the
CONSUMER and said so in its own commit message. This is the correction.

Three defects, all of them placement:

1. It sat in the engram. Ingest is a LANGUAGE concern — every el program
   touching any modality needs it, and the engram is merely one el program
   that happens to hold a graph. The geometry surface is now defined in
   el_runtime.c immediately ABOVE the engram section and depends on nothing
   inside it. Delete the entire engram and geometry still enters el.

2. It marshalled the vector as a hex STRING, because el had no first-class
   geometry value — which reintroduced text as the TRANSPORT medium one layer
   below the problem being fixed. Geometry is now an el value: a magic-tagged
   heap object carried in el_val_t, same discipline as List/Map. Hex survives
   only as an adapter at the edge, which is all an encoding should ever be.

3. It needed an arbitrary `dim <= 8192` bound purely to size an allocation
   from a caller's CLAIM about a string's length. A value carries its own
   width, so the width is derived and never asserted. The bound is gone, not
   raised — there is nothing left to validate.

Language surface, none of it engram-prefixed: geometry_new / _dim / _is /
_get / _set / _norm / _free, geometry_from_f32le_hex + geometry_to_f32le_hex
as the wire adapters, realizer_register(modality, fn_name), realizer_has, and
transduce(signal, modality) -> Geometry.

REALIZERS ARE DECLARABLE IN EL. This is the part that makes the move real
rather than nominal: registration resolves a name with dlsym against the
running binary, the identical mechanism http_set_handler already relies on,
because every el `fn name(...)` compiles to a global C symbol with that exact
name. So an ordinary el function IS a realizer and a new modality needs no
runtime patch. Verified end to end in lang/examples/transduce.el: an el-defined
tone_realizer is registered by name, transduce dispatches to it, and the
signal demonstrably reaches it (distinct signals produce distinct geometry).

A modality with no realizer transduces to NOTHING. There is deliberately no
built-in realizer, not even for text — silently embedding a description of a
signal and calling that perception is the exact defect this ends.

engram/src/server.el is migrated: POST /api/nodes decodes "emb" hex exactly
once, at the edge, into a Geometry, and everything below that line moves
geometry. The wire is unchanged because production clients speak it. "dim" is
now an ASSERTION about the vector, not the source of its width; disagreement
is a rejected ingest, not a silent reinterpretation.

#141's engram_node_set_emb becomes a DEPRECATED WRAPPER over
geometry_from_f32le_hex + node_attach_geometry — kept only because the runtime
ships as an SDK asset and a downstream binary may link the symbol. Its exact
contract, negative cases included, is preserved and re-verified.

ingest.el's `fn transduce` is renamed transduce_manifold. Mechanically it had
to yield the name (duplicate C symbol, a hard compile error, measured). But it
was never signal->geometry: it chunks already-extracted content into a node+edge
manifold, one layer up, and had taken the name belonging to the primitive
underneath it. Behaviour unchanged.

PROPERTIES FROM #141 PRESERVED, each re-measured on a scratch engram (:8971,
never prod :8742):
  - off-dimension vectors stored but NOT indexed — the HNSW build loop still
    filters on n->emb_dim == dim at four sites, so a 64-dim voice vector is
    durable and addressable without perturbing the 768-dim canonical index
  - geometry makes a node ineligible for embed_backfill: after backfill the
    64-dim voice node was still 64-dim while the text control acquired 768
  - the create response reports whether geometry landed, and the node document
    always emits emb_dim and embedded

Read-back with control and negatives, all verified against a PID-confirmed
fresh binary: geometry node emb_dim=64 embedded=true / emb_set=1; text-only
control emb_dim=0 embedded=false / emb_set=0; malformed hex, ragged length,
and dim-disagreement each emb_set=0.

Two compiler landmines found by reading the generated C rather than trusting a
successful build, both documented at their sites: elc lowers `a == b` to
str_eq unless both operand NAMES are in the per-function int-name set (which
does NOT propagate into nested if-expression blocks — the first cut would have
strcmp'd two integers as pointers on the first geometry-bearing request), and
`+` lowers to string concat when either operand is a user-defined call.

DESIGN.md

@@ -0,0 +1,630 @@
+# El Test Framework — Design
+
+**Status:** draft for review
+**Author:** Neuron
+**Date:** 2026-08-15
+**Worktree:** `/Users/will/Development/neuron-technologies/el-worktrees/elc-memory-investigation`
+
+---
+
+## 0. The forcing requirement
+
+We have a confirmed quadratic in `elc`. Peak memory in the old shipped binary and wall-clock in
+the current source both grow as O(input²). We cannot fix it, because we cannot test it.
+
+Everything in this document is downstream of one sentence: **a test framework must be able to fail
+a build when an operation's growth curve degrades from linear to quadratic.**
+
+That is not a nice-to-have bolted onto a correctness framework. It is the requirement that
+determines the architecture. Correctness testing is the easy half.
+
+Second-order requirement, learned the hard way tonight: **the framework must report per-test timing
+by default.** The current framework prints `N passed, M failed` and nothing else. That is why a
+3.58-second test file sat in the suite unnoticed. A framework that is structurally blind to time
+cannot surface the defect class we most need to catch.
+
+---
+
+## 1. What exists today, measured
+
+### 1.1 Two competing systems, neither complete
+
+**System A — `lang/runtime/test.el`.** Manual registration, El-level.
+
+**System B — the compiler's `test { }` block + `elc --test`.** Emits its own harness `main()`
+with `__el_pass` / `__el_fail` globals (`codegen.el:3777-3796`).
+
+They do not share a result model. Neither has timing. Both are in the tree.
+
+### 1.2 Specific defects in System A
+
+| Defect | Location | Consequence |
+|---|---|---|
+| All state as JSON strings in a global string-keyed map | `test.el` throughout | every assertion is `state_get` → `str_to_int` → `int_to_str` → `state_set` |
+| Failure list appended by string slice + concat | `_test_json_append` | O(n²) in failure count |
+| One OS thread spawned per test | `_test_run_one` via `__thread_create`/`__thread_join` | thread spawn per test, purely to get dispatch-by-name through dlsym |
+| Manual registration pairing a string to a function name | `test_case(name, fn_name)` | typo ⇒ test silently never runs, suite still reports pass |
+| Counters are assertion-level, global | `_test_pass_count` etc. | no per-test record exists at all |
+| No timing, no structured output, no fixtures, no tags, no filtering, no parameterization, no benchmarks | — | — |
+
+The registration defect is the serious one. It is not a slow framework, it is a framework that can
+report success for tests that did not execute.
+
+### 1.3 Measured cost structure
+
+Per test file, current build model:
+
+| Step | Time |
+|---|---|
+| `elc` compile `.el` → `.c` | 0.00s (small files) |
+| **`cc` el_runtime.c → .o** | **0.14s** |
+| `cc` test .c → .o | 0.02s |
+| link | 0.02s |
+
+> **STALE as of el #132 — re-measured 2026-08-16.** The `test_compiler` figure below was
+> *entirely* the `strlen`-per-character quadratic, now fixed. Re-measured on the same host:
+> **3.58s → 0.03s (119x)**, and the 422 KB compiler concatenation likewise compiles in 0.03s.
+> The table is retained only as the historical record that motivated the gate. The remaining
+> per-file cost is the redundant `el_runtime.c` rebuild, which §9's compile-once architecture
+> addresses.
+
+Per-file `elc` time across the existing suite:
+
+| File | Bytes | elc time |
+|---|---|---|
+| `test_compiler` | 29,685 (+394 KB of imports) | **3.58s** |
+| `string_test` | 18,545 | 0.01s |
+| all other 9 files | 2.2–10 KB | 0.00s |
+
+Two distinct defects in two distinct regimes:
+
+1. **`test_compiler.el` imports all five compiler sources** — 394 KB in one translation unit. Its
+   3.58s is entirely the quadratic. It is the only file where the quadratic bites.
+2. **Every other file's cost is 100% redundant `el_runtime.c` rebuilds** — 480 KB of identical C,
+   recompiled once per test file.
+
+Neither is fixed by making the compiler faster. Both are fixed by the architecture below, and the
+speedup is a by-product of building it correctly, not the goal.
+
+### 1.4 The asset worth keeping
+
+`codegen.el:3651-3652` already collects `test_names` / `test_c_names` — **the compiler already does
+compile-time test discovery.** It then discards that registry into a hardcoded `main()`.
+
+That registry is precisely the seam Go's `_testmain.go` and Rust's `test_main_static` are built on.
+The mechanism we need is half-built and wired to the wrong thing.
+
+---
+
+## 2. Grounding — the common spine of excellent frameworks
+
+Researched from primary sources: Go `testing`/`go test`, Rust `libtest`/Criterion, JUnit 5 Platform,
+NUnit 3, JMH, Google Benchmark. Six invariants hold across all of them.
+
+1. **A registry is built before execution** — `(name, metadata, fn-ptr)` triples. Go generates it
+   from an AST scan; Rust synthesizes it in a compiler pass; JMH emits it as a build-time resource;
+   JUnit/NUnit build it reflectively. **Reflection is an implementation of the registry on runtimes
+   where it is cheap. It is never the architecture.**
+
+2. **Discovery strictly precedes execution.** Every good capability — filtering, listing, counting,
+   sharding, IDE trees, re-run-failed-only, dry runs — is a consequence of this ordering.
+
+3. **A hierarchy with stable, path-shaped unique IDs.** `TestFoo/subcase_2`. Selection is regex over
+   that path, one pattern per level.
+
+4. **The framework is a prebuilt library; only the entry point is generated.** "Compile once, link
+   many" is always: framework archive compiled once + a small generated table + one
+   `MainStart(deps, registry)` call. Nobody recompiles the harness per test file.
+
+5. **Execution emits an event stream; reporters are downstream renderers.** Human text, NDJSON,
+   JUnit XML, TAP are all transforms of one event stream. Go's one architectural mistake is doing
+   this backwards — `test2json` parses human output, and has shipped bugs when user output contains
+   `--- PASS:`.
+
+6. **A dependency-injection seam at the boundary.** Go's `testdeps.TestDeps` exists so `testing`
+   can avoid importing `regexp`, profilers, and coverage. The execution core knows nothing about
+   output formats.
+
+---
+
+## 3. Architecture
+
+### 3.1 The seam
+
+```
+  ┌─────────────────────────────────────────────────────────────┐
+  │ user code:  foo.el  with  test { } / bench { }  blocks       │
+  └───────────────────────────┬─────────────────────────────────┘
+                              │  elc --test
+                              ▼
+  ┌─────────────────────────────────────────────────────────────┐
+  │ generated C (per suite, tiny):                               │
+  │   __el_test_fn_0 .. _N      lowered test/bench bodies        │
+  │   __el_registry[]           static table: name/kind/file/    │
+  │                             line/tags/sizes/expected-O       │
+  │   __el_dispatch(i)          generated switch → body          │
+  │   main() { return el_test_main(argc, argv); }                │
+  └───────────────────────────┬─────────────────────────────────┘
+                              │  cc + link  (registry only)
+                              ▼
+  ┌─────────────────────────────────────────────────────────────┐
+  │ libeltest.a  — PREBUILT ONCE                                 │
+  │   • el_runtime.o        (the 480 KB, compiled once, ever)    │
+  │   • eltest.o            the runner, WRITTEN IN EL            │
+  │       discovery view · filtering · execution · fixtures ·    │
+  │       timing · benchmark harness · curve fitting · reporters │
+  └─────────────────────────────────────────────────────────────┘
+```
+
+The framework is written in El, compiled to C once, archived. Per-suite compilation touches only
+the generated registry. This is Go's model, and it is strictly better for us than Go's because we
+own the compiler and already have the AST — no separate source-scanning pass is needed.
+
+### 3.2 Why the runner is in El and the registry is in C
+
+El has no closures and no first-class function pointers. The registry must therefore hold C function
+pointers, and it is generated C.
+
+The runner stays in El and reaches the registry through a small builtin surface — indices, not
+pointers:
+
+```
+__el_reg_count() -> Int
+__el_reg_name(i) -> String
+__el_reg_file(i) -> String
+__el_reg_line(i) -> Int
+__el_reg_kind(i) -> Int          // 0=test 1=bench
+__el_reg_tags(i) -> Int
+__el_reg_sizes(i) -> String      // JSON array, empty for tests
+__el_reg_expect(i) -> Int        // complexity class enum, 0 = none
+__el_reg_invoke(i) -> Int        // runs the body via the generated switch
+```
+
+Nine builtins. Everything else — filtering, lifecycle, statistics, curve fitting, all reporters —
+is El. That satisfies "written in El" without pretending El can do something it cannot.
+
+### 3.3 Result model
+
+The unit is a **result record**, not a counter:
+
+```
+TestResult {
+  id        String     // slash path: "parser/handles_empty_input/case_3"
+  file      String
+  line      Int
+  status    Status     // Pass | Fail | Error | Skip
+  duration  Int        // nanoseconds, ALWAYS populated
+  message   String     // assertion detail: expected vs actual
+  output    String     // captured stdout/stderr for this test
+  assertions Int
+}
+```
+
+`Fail` = an assertion failed. `Error` = unexpected crash/abort. This distinction is load-bearing —
+every CI consumer depends on it, and the JUnit XML schema encodes it as distinct elements.
+
+---
+
+## 4. Authoring surface
+
+### 4.1 Tests
+
+`test { }` already exists. Keep it. Add subtests and hierarchy:
+
+```el
+test "parser/empty input" {
+    assert_that(parse(""), is_err())
+}
+
+test "parser/table" {
+    for case in [["", 0], ["a", 1], ["a b", 2]] {
+        subtest(case[0]) {
+            assert_that(token_count(case[0]), equals(case[1]))
+        }
+    }
+}
+```
+
+Subtest IDs compose as `parser/table/a_b`. Filtering is `--run 'parser/table/.*'`, one regex per
+path segment, exactly as Go does.
+
+**We do not build a parameterized-test annotation system.** Table-driven loops plus subtests subsume
+`@ParameterizedTest`, `@MethodSource`, `@CsvSource`, and `TestCaseSource` entirely, at zero framework
+surface. This is Go's single biggest ergonomic win over JUnit and NUnit.
+
+### 4.2 Fixtures
+
+Per-file and per-test only, plus a LIFO cleanup stack:
+
+```el
+setup_all  { ... }      // once per suite
+setup      { ... }      // before each test
+teardown   { ... }      // after each test
+teardown_all { ... }
+```
+
+and inside a test, `cleanup { ... }` registering LIFO-ordered teardown.
+
+**We do not build JUnit 5's extension SPI** — seventeen callback interfaces, hierarchical stores,
+registration ordering rules. That complexity is the price of retrofitting a plugin ecosystem onto a
+twenty-year-old reflective framework. Go's `t.Cleanup` covers roughly 90% of what `@AfterEach` is
+used for at a fraction of the surface.
+
+### 4.3 Assertions — constraint model
+
+One entry point, composable constraint values (NUnit's model, which avoids the N² overload
+explosion):
+
+```el
+assert_that(actual, equals(expected))
+assert_that(xs,     has_length(3))
+assert_that(s,      contains("foo").and(starts_with("bar")))
+assert_that(f,      is_within(0.01).of(3.14))
+```
+
+A constraint is a value with `apply_to(actual) -> ConstraintResult`, and the result knows how to
+describe its own failure. Custom constraints are ordinary user types.
+
+**Every failure message must name file, line, the expression text, and both values.** We capture
+expression source text at compile time — we have the AST, so we can do this better than any
+runtime-introspection framework.
+
+Legacy `assert_true` / `assert_eq` / etc. stay as thin wrappers for migration.
+
+---
+
+## 5. Benchmarks
+
+### 5.1 The loop
+
+Adopt `b.Loop()`, not `b.N`. Go spent fifteen years on `b.N` before concluding `b.Loop` was right;
+we skip that.
+
+```el
+bench "str_concat" {
+    let s = make_input(bench_n())
+    for bench_loop() {
+        black_box(str_concat(s, "x"))
+    }
+}
+```
+
+Three properties that make this the correct choice for a C target:
+
+1. **The timer auto-resets on first call**, so setup above the loop is excluded *by construction*
+   rather than by the author remembering `ResetTimer`.
+2. **`N` is hidden**, so it cannot be misused.
+3. **The harness owns the loop shape**, which lets us insert an optimization barrier the C compiler
+   cannot see through. `black_box(v)` lowers to `asm volatile("" :: "r"(&v) : "memory")`. Since we
+   emit a single translation unit, dead-code elimination of a benchmark body is a live hazard —
+   this is our version of JMH's `Blackhole` problem, solved in the harness rather than delegated to
+   the user.
+
+### 5.2 Iteration scaling
+
+Use Go's `predictN` heuristics verbatim. They are battle-tested and cheap:
+
+```
+n = goal_ns * prev_iters / prev_ns    // multiply before divide — precision on sub-ns ops
+n += n / 5                            // 20% headroom, overshoot rather than re-loop
+n = min(n, 100 * last)                // never grow more than 100× per step
+n = max(n, last + 1)                  // guarantee forward progress
+n = min(n, 1_000_000_000)             // hard ceiling
+```
+
+Report `n` rounded to 1/2/3/5 × 10ᵏ so runs are comparable.
+
+### 5.3 Sampling
+
+Criterion's shape, because it is correct near timer resolution:
+
+- **Warmup**: iteration counts 1, 2, 4, 8… until cumulative time exceeds the warmup budget.
+- **Measurement**: collect `sample_size` samples at iteration counts `[d, 2d, 3d, …, Nd]`.
+- **Estimate**: slope of a linear regression of iteration-count vs elapsed time. The intercept
+  absorbs fixed overhead.
+- **Time whole samples, never individual iterations.** This is the single most important detail —
+  it defeats timer-resolution error on nanosecond operations.
+
+Outliers classified by modified Tukey (±1.5 IQR mild, ±3 IQR severe), **reported but retained**.
+
+---
+
+## 6. Complexity gating — the centerpiece
+
+This is the part that makes the quadratic fixable, and the part nobody in the mainstream has
+finished. Google Benchmark's `Complexity()` fits the curve and *reports* it. We declare it and
+**gate** on it.
+
+### 6.1 Surface
+
+```el
+bench "elc_compile" over n in [16, 32, 64, 128, 256, 512, 1024] expect O(n) {
+    let src = synth_source(bench_n())
+    for bench_loop() { black_box(compile(src)) }
+}
+```
+
+Alternative with no new syntax, if the parser change is judged too invasive — `bench_sizes([...])`
+and `bench_expect("O(n)")` as calls inside the block. **Recommendation: declarative.** Runtime calls
+mean `--list` cannot show the invariant without executing, which breaks the discovery-precedes-
+execution invariant from §2.
+
+### 6.2 Fitting
+
+Per Google Benchmark `src/complexity.cc`. For candidate curves
+`{O(1), O(log n), O(n), O(n log n), O(n²), O(n³)}`, one-parameter least squares, no intercept:
+
+```
+coef = Σ(tᵢ · gᵢ) / Σ(gᵢ²)
+rms  = sqrt( Σ(tᵢ − coef·gᵢ)² / k ) / mean(t)     // normalized
+```
+
+Best fit = lowest normalized RMS. User-supplied lambda curves also supported.
+
+### 6.3 Gate logic
+
+1. **FAIL** if the best-fit curve is strictly worse than declared, ordering
+   `O(1) < O(log n) < O(n) < O(n log n) < O(n²) < O(n³)`. Print the fitted coefficient and the full
+   per-size table.
+2. **FAIL** if the declared curve's normalized RMS exceeds a threshold (start at 0.10). This catches
+   the case where *no* candidate fits — noise, a cache cliff, or a phase change. Report
+   `INDETERMINATE` honestly rather than gating on garbage.
+3. **WARN** if the best fit is strictly better than declared — either an optimization landed and the
+   annotation should tighten, or the sweep is too narrow to expose real behaviour.
+4. **REFUSE to gate** on fewer than 5 distinct sizes spanning under 2 decades, geometrically spaced.
+   Say so loudly rather than producing a meaningless fit.
+
+### 6.4 Why gate on the exponent, not wall-clock
+
+- **Machine-independent.** The fitted exponent is a property of the algorithm; the coefficient is a
+  property of the machine. Gating on the exponent makes CI hardware heterogeneity, noisy neighbours,
+  and thermal throttling irrelevant — they scale `coef`, not `g`.
+- **No stored baseline.** No artifact storage, no golden-file drift. The invariant lives in the
+  source next to the code and is reviewed in the same PR.
+- **It catches the failure mode that actually ships.** An O(n) lookup inside an O(n) loop is
+  invisible at n=100 in a unit test and catastrophic at n=100,000 in production. Constant-factor
+  regressions are annoying. Complexity regressions are outages. Ours was a 27 GB outage.
+
+### 6.5 The deterministic gate — the one that would have caught us
+
+Wall-clock needs statistics. **Allocation counts do not.** They are perfectly deterministic.
+
+> **Correction, 2026-08-16 — count alone is NOT sufficient. Gate on BOTH count and bytes.**
+>
+> Measured against two El programs, one allocating once per item and one rebuilding its
+> accumulator each iteration:
+>
+> | n | linear allocs / bytes | quadratic allocs / bytes |
+> |---|---|---|
+> | 100 | 100 / 290 | 100 / 5,150 |
+> | 200 | 200 / 690 | 200 / 20,300 |
+> | 400 | 400 / 1,490 | 400 / 80,600 |
+> | 800 | 800 / 3,090 | 800 / 321,200 |
+>
+> The quadratic program's allocation **count is exactly linear** — 100/200/400/800, identical to
+> the healthy program. A count-only gate passes it clean. **Bytes** catch it: each doubling of n
+> quadruples bytes (ratios 3.94, 3.97, 3.99 → 4.0 = O(n²)) where the linear program converges
+> on 2.0.
+>
+> This is precisely elc's own defect shape — a copy-on-write accumulator reallocating once per
+> pass (count linear) into a proportionally larger buffer (bytes quadratic).
+>
+> Therefore `expect allocs O(n)` **fits count and bytes independently and fails if EITHER exceeds
+> the declared curve**, reporting which signal broke. "count linear, bytes quadratic" is a precise,
+> directly actionable diagnosis.
+>
+> **`el_peak_rss()` is CONTEXT ONLY — never gate on it.** It is perturbed by the allocator and by
+> the page cache. Allocation volume is the invariant; RSS and malloc/free churn are merely the two
+> surfaces it shows on. The old shipped compiler paid the same quadratic in RSS that the rebuilt
+> one pays in churn.
+>
+> **Measure rate, not level.** A guard reading swap *level* saw 97% on a thrashing host and 97% on
+> a healthy one; only *rate* separated them. A growth exponent is a rate; a single measurement is
+> a level. That is why the gate fits a curve across a sweep instead of comparing one number to a
+> threshold.
+
+> **Second correction, same day — THE ALLOCATION GATE ALONE WOULD HAVE MISSED THE REAL BUG.**
+>
+> el #132 found the actual elc quadratic: `strlen()` called inside `str_char_code()` and
+> `str_slice()`, so the lexer rescanned the remaining input on every character. Pure CPU.
+> **Zero allocation.** `str_char_code` is a bounds check and an index — it allocates nothing.
+>
+> Measured on three controlled specimens (`lang/.work/fitprobe.el`), growth ratio per doubling of
+> n across n = 200/400/800/1600:
+>
+> | specimen | allocs | bytes | time | what it proves |
+> |---|---|---|---|---|
+> | `linear` — one alloc per item | 2.00 2.00 2.00 → **O(n)** | 2.16 2.07 2.23 → **O(n)** | 0.83 2.00 2.05 → **O(n)** | clean baseline |
+> | `accum` — rebuilds accumulator | 2.00 2.00 2.00 → **O(n)** | 3.97 3.99 3.99 → **O(n²)** | noisy | count misses, **bytes catches** |
+> | `compute` — n scans over n chars | 0 → **FLAT** | 0 → **FLAT** | 3.93 4.01 3.96 → **O(n²)** | **both alloc signals blind; only time catches** |
+>
+> `compute` is el #132's shape exactly. A gate fitting only allocation count and bytes classifies
+> it as FLAT and passes it. **The gate as originally specified would not have caught the defect it
+> was created for.**
+>
+> Therefore the gate fits **THREE** signals and fails if ANY exceeds its declared curve:
+>
+> ```
+> bench "elc_compile" over n in [...] expect time O(n) allocs O(n) bytes O(n) { ... }
+> ```
+>
+> - **allocs (count)** — deterministic, zero-noise. Catches per-item allocation growth.
+> - **allocs (bytes)** — deterministic, zero-noise. Catches accumulator-rebuild quadratics that
+>   count cannot see.
+> - **time** — noisy, needs the sweep and statistics. The ONLY signal that sees pure-compute
+>   complexity regressions. Gate on the fitted *exponent*, never on absolute duration, so CI
+>   hardware variance scales the coefficient and leaves the classification intact.
+>
+> The deterministic signals remain preferable where they apply — they need no statistics and are
+> correct on the first run. They are simply not sufficient.
+>
+> **`black_box` is mandatory, and consuming the result is NOT enough.** The first version of
+> `compute` accumulated `total + 1` in a nested loop and reported **0 µs at every n** while
+> returning a numerically correct n². Clang recognised the idiom and closed the loop to a
+> multiply. Feeding the result into output did not prevent it. Only making the inner operation an
+> opaque external call restored the real curve. A benchmark harness that trusts the user to defeat
+> the optimiser will silently measure nothing — and report success while doing it.
+
+Instrument the runtime with allocation counters and fit *those* against n instead of time:
+
+```el
+bench "elc_compile" over n in [...] expect O(n) allocs O(n) { ... }
+```
+
+Zero noise, zero statistics, always gateable, correct on the first run on any machine. Go reports
+`allocs/op` and `B/op`; **nobody fits them against n.** That is an open opportunity and it is exactly
+our bug: elc's defect is quadratic *allocation volume*, which the old binary paid in RSS and the
+current source pays in malloc/free churn.
+
+An `expect allocs O(n)` assertion on `elc`'s compile path would have failed the build the day the
+quadratic was introduced.
+
+Required runtime additions: `__el_alloc_count()`, `__el_alloc_bytes()`, `__el_peak_rss()`.
+
+### 6.6 Constant-factor gate (secondary, opt-in)
+
+Mann-Whitney U at α = 0.05, noise floor 1%, medians with 95% CIs, `~` for not-significant. Requires
+`--count >= 9`. Off by default on CI; opt-in per benchmark.
+
+**Exit nonzero on regression.** Both benchstat and Criterion always exit 0, which is why every shop
+using them wrote a wrapper. We do not repeat that omission.
+
+---
+
+## 7. Output
+
+**Structured events are the source of truth.** Human text is rendered from them. We do not repeat
+Go's parse-the-human-output design.
+
+Event stream, NDJSON, one object per line, streamed live:
+
+```json
+{"time":"...","action":"run","test":"parser/empty"}
+{"time":"...","action":"output","test":"parser/empty","output":"..."}
+{"time":"...","action":"pass","test":"parser/empty","elapsed":0.0031}
+{"time":"...","action":"bench","test":"str_concat","n":1024,"ns_op":41.2,"allocs_op":3,"bigo":"N","rms":0.03}
+```
+
+Renderers, all downstream and pluggable:
+
+| Format | Flag | Use |
+|---|---|---|
+| Human | default | terminal, **per-test duration always shown** |
+| NDJSON | `--json` | tooling, history, flaky detection |
+| JUnit XML | `--junit-xml=PATH` | every CI system on earth |
+| TAP | `--tap` | optional |
+
+JUnit XML per the de-facto schema: `testsuites` → `testsuite` → `testcase`, with `time` in seconds
+as a decimal, `file`/`line` attributes, and `failure` vs `error` vs `skipped` as distinct child
+elements. Absence of a child element means pass. Emit `<testsuites>` even for a single suite, and
+parse both shapes on input.
+
+---
+
+## 8. CLI
+
+```
+--list                     print the registry, run nothing
+--list-json                machine-readable registry
+--run PATTERN              slash-separated regex per path segment
+--tag EXPR                 tag expression: fast & !slow
+--shard I/N                deterministic sharding for CI parallelism
+--count N                  repetitions, for statistics
+--bench PATTERN            run benchmarks (off by default in test runs)
+--benchtime DUR            per-benchmark time budget
+--junit-xml PATH
+--json
+--isolate                  re-exec per test on crash, so one SIGSEGV doesn't lose the run
+--timeout DUR
+--fail-fast
+```
+
+`--list` / `--list-json` / `--shard` cost roughly thirty lines because the registry already exists
+before `main` does anything. That is the dividend of discovery-precedes-execution.
+
+---
+
+## 9. Build model
+
+```
+# once, ever (or when the runtime/framework changes):
+cc -c el_runtime.c        -o el_runtime.o
+elc eltest.el > eltest.c && cc -c eltest.c -o eltest.o
+ar rcs libeltest.a el_runtime.o eltest.o
+
+# per suite:
+elc --test foo_test.el > foo_test.c     # registry + bodies only
+cc foo_test.c libeltest.a -o foo_test
+```
+
+The 0.14s × N of redundant runtime rebuilds disappears — not because we optimized it, but because
+one-runner-over-many-suites requires compile-once-link-many as a structural precondition.
+
+---
+
+## 10. Bootstrap and self-hosting
+
+The framework's own tests are `test { }` blocks run by the framework. Same fixpoint discipline the
+compiler already applies to itself.
+
+1. Build the framework using the *existing* harness for its first tests (stage 0).
+2. Rebuild the framework's tests as `test { }` blocks run by the new runner (stage 1).
+3. Verify stage 1 reports identical results to stage 0.
+4. From then on, the framework is tested by itself.
+
+A framework that cannot run its own suite is not evidence of anything. This is a correctness proof,
+not a claim.
+
+---
+
+## 11. Explicitly not building
+
+| Rejected | Why |
+|---|---|
+| Naming-convention discovery (`fn test_foo`) | `test { }` is a real declaration. Go's `TestXxx` exists only because Go had no better hook — and it needs a heuristic to avoid matching `TesticularCancer`. |
+| Reflection or symbol-table scanning | Slow, fragile under LTO/strip/dead-strip, and unnecessary when we own the compiler. |
+| Parsing human output into structure | Go's `test2json` is its one clear architectural mistake. |
+| JUnit 5's extension SPI | Seventeen callback interfaces to retrofit plugins onto a reflective framework. Not our problem. |
+| `@ParameterizedTest` machinery | Table-driven loops + subtests subsume it at zero surface. |
+| NUnit's out-of-process agents | They bridge CLR versions and AppDomains. We emit one native binary. Keep `--isolate` as crash fallback only. |
+| JMH-style forking by default | Forks exist because JIT profiles are per-process. AOT C has no such state. Keep `--fork` available, not default. |
+| Exit 0 on regression | benchstat and Criterion both do this, and every user writes a wrapper. |
+| Dynamic runtime test registration | Breaks `--list`, sharding, and individual selection. Registry stays static. |
+
+---
+
+## 12. Phasing
+
+| Phase | Content | Gate |
+|---|---|---|
+| **1** | Registry emission in codegen; 9 builtins; `el_test_main` skeleton in El; result records; per-test timing; human + NDJSON output | existing 11 test files pass, with timing |
+| **2** | `libeltest.a` build model; subtests; filtering; `--list`; fixtures; constraint assertions; JUnit XML | suite runs in one binary; runtime compiled once |
+| **3** | `bench { }`, `bench_loop`, `black_box`, `predictN`, Criterion sampling | benchmarks produce stable ns/op |
+| **4** | Allocation counters; complexity fitting; `expect O(...)` gate | **an `expect allocs O(n)` benchmark on `elc` fails on the current quadratic** |
+| **5** | Migrate both legacy systems; delete `runtime/test.el`; self-host | framework runs its own suite |
+
+Phase 4 is the deliverable that matters. Phases 1–3 exist to make it possible.
+
+---
+
+## 13. Open questions for review
+
+1. **Declarative `over n in [...] expect O(...)` syntax vs runtime calls.** I recommend declarative
+   (§6.1) so `--list` can show invariants without executing. It costs parser work. Your call.
+2. **`bench { }` as a new block form** — parallel to `test { }`, or a modifier on it?
+3. **Scope of the constraint model.** Full composable constraints, or start with a flat assertion set
+   and add constraints later? Full model is more surface but avoids a second migration.
+4. **Does `runtime/test.el` get deleted or kept as a deprecated shim?** I lean delete — two systems
+   is how we got here.
+5. **Where does `libeltest.a` live** in the tree, and does `epm` need to know about it?
+6. **Allocation counters in `el_seed.c` or `el_runtime.c`?** AGENTS.md says `el_seed.c` is the sole
+   C dependency and hand-maintained; counters are OS-boundary-adjacent but not OS calls.
+7. **Is per-test timing enough, or do we want per-*assertion* timing** for finding slow helpers?
+
+---
+
+## 14. What this document is not
+
+This is a design, not a measurement. Every performance claim about the *current* system in §1 is
+measured and reproducible in this worktree. Every claim about the *proposed* system is a prediction.
+None of it is verified until Phase 1 runs and Phase 4 fails a build on the real quadratic.

engram/src/server.el

@@ -41,17 +41,29 @@ fn strip_query(path: String) -> String {
     str_slice(path, 0, q)
 }
 
+// query_param — extract one query-string value, URL-DECODED.
+//
+// The decode step was missing (found 2026-08-15): a claim sent as
+// "test%20claim" arrived at engram_assert_json still percent-encoded and was
+// stored/compared that way, so any value containing a space, &, =, or non-ASCII
+// character silently became a different string than the caller sent. Affects
+// every GET route that reads params this way, not just /api/assert.
 fn query_param(path: String, key: String) -> String {
     let q: Int = str_index_of(path, "?")
     if q < 0 { return "" }
     let qs: String = str_slice(path, q + 1, str_len(path))
-    let needle: String = key + "="
-    let pos: Int = str_index_of(qs, needle)
+    // Anchor the match to a real key boundary: prefixing "&" and searching for
+    // "&key=" means "q" can never match inside "faq=". (Found 2026-08-15:
+    // "?faq=X&q=Y" returned X for key "q" — a silently wrong value, not an
+    // error.) The leading "&" makes the first parameter match the same way.
+    let hay: String = "&" + qs
+    let needle: String = "&" + key + "="
+    let pos: Int = str_index_of(hay, needle)
     if pos < 0 { return "" }
-    let after: String = str_slice(qs, pos + str_len(needle), str_len(qs))
+    let after: String = str_slice(hay, pos + str_len(needle), str_len(hay))
     let amp: Int = str_index_of(after, "&")
-    if amp < 0 { return after }
-    str_slice(after, 0, amp)
+    let raw: String = if amp < 0 { after } else { str_slice(after, 0, amp) }
+    return __url_decode(raw)
 }
 
 fn query_int(path: String, key: String, default_val: Int) -> Int {
@@ -235,6 +247,24 @@ fn persist_bulk() -> Int {
     return persist_canonical()
 }
 
+// COMPILER LANDMINE, measured 2026-08-16 — do not inline this back into the
+// caller. elc lowers `a == b` to numeric comparison only when both operand
+// NAMES are in the per-function int-name set, which `let x: Int` populates.
+// That registration does NOT propagate into a nested if-expression block: the
+// first cut of the geometry-ingest path wrote `let claimed: Int = ...` and
+// `let got: Int = ...` inside the else-arm and `claimed == got` came out of
+// codegen as `str_eq(claimed, got)` — strcmp on two integers reinterpreted as
+// pointers, i.e. a segfault on the first geometry-bearing request. Read back
+// out of the generated C, not guessed. Function PARAMETERS annotated `: Int`
+// do register reliably (verified: `if (claimed == actual)`), so the comparison
+// lives in a function of its own. Note also the explicit `return`s — a trailing
+// if-EXPRESSION at a function tail emits as a statement and the function
+// returns 0 regardless, which is the same probe's second finding.
+fn width_agrees(claimed: Int, actual: Int) -> Int {
+    if claimed == actual { return 1 }
+    return 0
+}
+
 // INCOMPLETE-ROUTE FIX (2026-07-24 self-review): this route silently dropped
 // label, importance, tier, and tags — engram_node() defaults label to content
 // and importance to 0.5, so every node created over HTTP lost its metadata.
@@ -276,6 +306,45 @@ fn route_create_node(method: String, path: String, body: String) -> String {
         salience, importance, confidence,
         tier, tags
     )
+    // GEOMETRY INGEST — geometry-valued end to end (2026-08-16).
+    //
+    // The defect this route originally had: it accepted an "emb" field,
+    // returned 200 with a fresh id, and stored NOTHING, because engram_node_full
+    // has no vector parameter. The consequence was structural, not cosmetic —
+    // text was the only entry medium, so any non-text modality had to be
+    // DESCRIBED in prose, and what we then reasoned over was the geometry of the
+    // description, not of the signal.
+    //
+    // #141 fixed the drop but marshalled the vector as a hex STRING through
+    // engram_node_set_emb, which put text back as the TRANSPORT medium one layer
+    // below the problem being fixed. This is that correction: hex is decoded
+    // exactly ONCE, here at the edge, into a first-class Geometry, and every
+    // step below this line moves geometry rather than text. An encoding at the
+    // boundary is what an encoding is for.
+    //
+    // The WIRE is deliberately unchanged — "emb" is still little-endian float32
+    // hex (8 chars per component), the encoding the perception vessel's
+    // /voice/embed already emits — because production clients speak it. What
+    // changed is underneath it.
+    //
+    // "dim" is now treated as an ASSERTION about the vector the caller sent, not
+    // as the source of its width: a Geometry carries its own width. A stated dim
+    // that disagrees is a REJECTED ingest, not a silent reinterpretation. Omitting
+    // "dim" is fine and means "trust the vector", which is the honest default.
+    //
+    // Off-dimension vectors remain stored but not inserted into the resident HNSW
+    // index (its build loop filters on emb_dim), so a 64-dim voice geometry is
+    // durable and addressable without perturbing the 768-dim canonical index.
+    let emb_hex: String = json_get_string(body, "emb")
+    let emb_set: Int = if str_eq(emb_hex, "") { 0 } else {
+        let g: Geometry = geometry_from_f32le_hex(emb_hex)
+        let got: Int = geometry_dim(g)
+        let dim_raw: String = json_get_raw(body, "dim")
+        let claimed: Int = if str_eq(dim_raw, "") { got } else { json_get_int(body, "dim") }
+        let landed: Int = if width_agrees(claimed, got) > 0 { node_attach_geometry(id, g) } else { 0 }
+        let freed: Int = geometry_free(g)
+        landed
+    }
     let saved: Int = persist_node(id)
     // ORPHAN PREVENTION (ENGRAM_AUTOCONNECT): connect the fresh node to its
     // nearest embedded neighbors so it never enters the graph edgeless.
@@ -286,7 +355,11 @@ fn route_create_node(method: String, path: String, body: String) -> String {
         if added > 0 { let sv2: Int = persist_edges_since(ec0) }
         added
     } else { 0 }
-    "{\"id\":\"" + id + "\",\"content\":\"" + content + "\",\"node_type\":\"" + node_type + "\",\"connected\":" + int_to_str(connected) + "}"
+    // Report whether the supplied geometry actually landed. The old response
+    // was success-shaped regardless — 200 with an id while the vector was
+    // discarded — which is how the drop went unnoticed. A caller can now
+    // assert on emb_set instead of trusting the status code.
+    "{\"id\":\"" + id + "\",\"content\":\"" + content + "\",\"node_type\":\"" + node_type + "\",\"connected\":" + int_to_str(connected) + ",\"emb_set\":" + int_to_str(emb_set) + "}"
 }
 
 fn route_get_node(method: String, path: String, body: String) -> String {

engram/test/run_vindex_concurrency_tests.sh

@@ -0,0 +1,107 @@
+#!/usr/bin/env bash
+# run_vindex_concurrency_tests.sh — regression harness for the 2026-08-16 soul crash.
+#
+# Four halves. The SET is the point: it separates two hazards the original two-half
+# version conflated, and which have fixes in different files.
+#
+#   1. single          ASan+UBSan, one thread. MUST be clean. Hard failure.
+#
+#   2. readers         TSan, N readers, NO writer. Hazard (a): the visited set used
+#                      to live on the index, so two pure READS stamped each other's
+#                      epoch. Fixed in engram_vindex.c (frame-owned VVisit +
+#                      `const VIndex*` search). MUST be clean. Hard failure.
+#
+#   3. unsynchronized  TSan, writer + reader on a BARE index. Hazard (b): in-place
+#                      HNSW insert rewires existing elements' neighbour lists and
+#                      reallocs elems[]. EXPECTED TO RACE, PERMANENTLY. This is not
+#                      a bug to fix inside engram_vindex.c — it is the executable
+#                      proof that a publication boundary must exist above it.
+#                      Not a failure. If it ever goes CLEAN, the test stopped
+#                      interleaving and half 4 is no longer meaningful either.
+#
+#   4. published       TSan, owner + N readers through a publication boundary
+#                      (rwlock: readers shared, owner exclusive) mirroring
+#                      eg_vindex_view / eg_vindex_maintain in lang/runtime/el_runtime.c.
+#                      MUST be clean, and all inserts must land. Hard failure.
+#
+# See test_vindex_concurrency.c for the full story (SIGSEGV at ASCII address
+# "gramNode", heap corruption in xzm_realloc, etc).
+#
+# usage: run_vindex_concurrency_tests.sh
+set -uo pipefail
+
+HERE="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+RUNTIME="$(cd "$HERE/../../lang/runtime" && pwd)"
+WORK="$(mktemp -d)"
+trap 'rm -rf "$WORK"' EXIT
+
+SRC="$HERE/test_vindex_concurrency.c"
+VINDEX="$RUNTIME/engram_vindex.c"
+
+fail=0
+
+echo "== [1/4] single-threaded control under AddressSanitizer =="
+cc -std=c11 -g -O1 -fsanitize=address,undefined -fno-omit-frame-pointer \
+   -I"$RUNTIME" -o "$WORK/single" "$SRC" "$VINDEX" -lm || { echo "BUILD FAILED"; exit 2; }
+if ASAN_OPTIONS=detect_leaks=0 "$WORK/single" single; then
+  echo "   -> OK"
+else
+  echo "   -> FAIL: the single-threaded control must always be clean."
+  echo "      If this fails the bug is NOT (only) concurrency — look for a real"
+  echo "      out-of-bounds or lifetime error in engram_vindex.c."
+  fail=1
+fi
+
+cc -std=c11 -g -O1 -fsanitize=thread -fno-omit-frame-pointer \
+   -I"$RUNTIME" -o "$WORK/conc" "$SRC" "$VINDEX" -lm || { echo "BUILD FAILED"; exit 2; }
+
+# run_tsan <mode> <logfile>; echoes nothing, sets $tsan_raced
+run_tsan() {
+  TSAN_OPTIONS="halt_on_error=0" "$WORK/conc" "$1" >"$2" 2>&1
+  tsan_rc=$?
+  if grep -q "ThreadSanitizer: data race" "$2"; then tsan_raced=1; else tsan_raced=0; fi
+}
+
+echo
+echo "== [2/4] concurrent READERS, no writer (visited-set gate) =="
+run_tsan readers "$WORK/readers.log"
+if [ "$tsan_raced" = "1" ]; then
+  echo "   -> REGRESSION: two concurrent reads still race."
+  grep -m1 -A6 "ThreadSanitizer: data race" "$WORK/readers.log" | sed 's/^/      /'
+  echo "      The visited set was supposed to be owned by the call frame."
+  fail=1
+else
+  echo "   -> clean (concurrent reads are safe)"
+fi
+
+echo
+echo "== [3/4] writer+reader on a BARE index (expected-race probe) =="
+run_tsan unsynchronized "$WORK/unsync.log"
+if [ "$tsan_raced" = "1" ]; then
+  echo "   -> RACE DETECTED, as expected:"
+  grep -m1 -A4 "ThreadSanitizer: data race" "$WORK/unsync.log" | sed 's/^/      /'
+  echo "      In-place HNSW insert mutates existing elements. Not fixable inside"
+  echo "      engram_vindex.c — this is why the publication boundary exists."
+else
+  echo "   -> NOTE: no race reported. The probe did not interleave; half 4's"
+  echo "      clean result proves less than it should. Investigate."
+fi
+
+echo
+echo "== [4/4] owner+readers through the publication boundary (boundary gate) =="
+run_tsan published "$WORK/pub.log"
+if [ "$tsan_raced" = "1" ]; then
+  echo "   -> REGRESSION: the publication boundary did not serialize the owner."
+  grep -m1 -A6 "ThreadSanitizer: data race" "$WORK/pub.log" | sed 's/^/      /'
+  fail=1
+elif [ "$tsan_rc" != "0" ]; then
+  echo "   -> FAIL: boundary clean under TSan but the run failed:"
+  tail -3 "$WORK/pub.log" | sed 's/^/      /'
+  fail=1
+else
+  echo "   -> clean (readers project concurrently; the owner's inserts all landed)"
+fi
+
+echo
+[ "$fail" -eq 0 ] && echo "RESULT: PASS" || echo "RESULT: FAIL"
+exit "$fail"

engram/test/test_vindex_concurrency.c

@@ -0,0 +1,251 @@
+/* test_vindex_concurrency.c — regression test for the 2026-08-16 soul crash.
+ *
+ * WHAT BROKE: the soul daemon crash-looped (5 crashes in ~100s) with SIGSEGV in
+ * search_layer <- vindex_insert <- eg_vindex_sync, a SIGABRT, and a fault inside
+ * xzm_realloc's own freelist — i.e. heap corruption. The SIGSEGV address
+ * 0x65646f4e6d617267 is little-endian ASCII "gramNode": string bytes being
+ * dereferenced as an Elem vector pointer.
+ *
+ * ROOT CAUSE: VIndex owns its traversal scratch (visited[] + visit_epoch), and
+ * search_layer mutates it via visited_reset(). So the index is unsafe for ANY
+ * concurrent use — including two concurrent READS. soul.el starts http_serve_async
+ * (a thread per connection) and then runs awareness_run() on the main thread, which
+ * reaches the same global index through engram_activate; nothing serialized them.
+ *
+ * Neither hnswlib nor FAISS puts the visited set on the index: hnswlib checks one
+ * out of a VisitedListPool per query, FAISS uses a thread_local VisitedTable.
+ *
+ * THE ORIGINAL `concurrent` HALF CONFLATED TWO DISTINCT HAZARDS (2026-08-16). It ran
+ * a writer against a reader on one bare index, so it could not tell apart:
+ *
+ *   (a) READ/READ corruption — two searches stamping each other's visited epoch.
+ *       A defect INSIDE engram_vindex.c, fixable there, and now fixed: the visited
+ *       set moved to the call frame and vindex_search takes a `const VIndex*`.
+ *
+ *   (b) WRITE/READ corruption — vindex_insert rewires the neighbour lists of
+ *       EXISTING elements and reallocs elems[], so an insert is a mutation of the
+ *       whole structure. This is NOT fixable inside engram_vindex.c at any price:
+ *       it is inherent to in-place HNSW. It requires a publication boundary ABOVE
+ *       the data structure (el_runtime.c: eg_vindex_view / eg_vindex_maintain).
+ *
+ * Conflating them made the suite unfailable-then-unpassable: fixing (a) left (b)
+ * still racing, which reads as "the fix did not work" when in fact a different,
+ * correctly-located fix is what (b) needs. So the halves are now separate:
+ *
+ *   single          N clustered vectors, ONE thread, ASan. The CONTROL. Must always
+ *                   be clean. When this passes and a concurrent half fails, the defect
+ *                   is concurrency, not an out-of-bounds/logic error in the graph code.
+ *                   (On 2026-08-16 this control cleared all 13,820 real dim-768 store
+ *                   vectors under ASan, which DISPROVED an inspection-derived hypothesis
+ *                   about an out-of-bounds reverse-link write at engram_vindex.c:340.)
+ *
+ *   readers         N reader threads, NO writer, one shared index, TSan. This is
+ *                   hazard (a) in isolation. It RACED before the visited set moved off
+ *                   the index struct and must be CLEAN now. Hard gate.
+ *
+ *   unsynchronized  writer + reader on a bare index, TSan. Hazard (b) in isolation.
+ *                   EXPECTED TO RACE, permanently — it is the executable proof that
+ *                   the index cannot be made safe from the inside, and therefore that
+ *                   the publication boundary in el_runtime.c has to exist. If this
+ *                   ever goes clean, the test stopped interleaving; do not celebrate.
+ *
+ *   published       writer + readers through a publication boundary that mirrors
+ *                   eg_vindex_view / eg_vindex_maintain (rwlock: readers shared,
+ *                   the single owner exclusive), TSan. Must be CLEAN. Hard gate.
+ *                   This is what proves the shape of the runtime fix, in the same
+ *                   process, rather than asserting it.
+ *
+ * Absence of a crash does NOT mean absence of a race — always read the sanitizer
+ * verdict, never just the exit code.
+ *
+ * Build/run: engram/test/run_vindex_concurrency_tests.sh
+ */
+#include "engram_vindex.h"
+
+#include <pthread.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <stdint.h>
+
+#define DIM   128
+#define NVEC  3000
+#define SEED_N 50
+
+static VIndex* g_ix;
+static float*  g_vecs;
+
+/* Deterministic filler. Real embeddings are strongly correlated, not uniform noise;
+ * clustering keeps many candidates near-equidistant, which exercises the diversity
+ * heuristic and the visited set far harder than random vectors do. */
+static void fill_vectors(void) {
+    g_vecs = (float*)malloc((size_t)NVEC * DIM * sizeof(float));
+    if (!g_vecs) { fprintf(stderr, "OOM\n"); exit(1); }
+    for (int i = 0; i < NVEC; i++) {
+        int cluster = i % 8;
+        for (int d = 0; d < DIM; d++)
+            g_vecs[(size_t)i * DIM + d] =
+                (float)(((d + cluster * 7) % 13) / 13.0) +
+                (float)(((i * 2654435761u + (unsigned)d) % 97) / 9700.0);
+    }
+}
+
+static void* writer_fn(void* arg) {
+    (void)arg;
+    for (int i = SEED_N; i < NVEC; i++)
+        (void)vindex_insert(g_ix, (uint64_t)i, g_vecs + (size_t)i * DIM);
+    return NULL;
+}
+
+static void* reader_fn(void* arg) {
+    (void)arg;
+    uint64_t ids[8]; float ds[8];
+    for (int i = 0; i < 20000; i++)
+        (void)vindex_search(g_ix, g_vecs + (size_t)(i % NVEC) * DIM, 8, 0, ids, ds);
+    return NULL;
+}
+
+static int run_single(void) {
+    printf("[single] inserting %d vectors on one thread (ASan control)\n", NVEC);
+    g_ix = vindex_create(DIM, 0, 0);
+    if (!g_ix) { fprintf(stderr, "[single] vindex_create failed\n"); return 1; }
+    for (int i = 0; i < NVEC; i++) {
+        if (vindex_insert(g_ix, (uint64_t)i, g_vecs + (size_t)i * DIM) != 0) {
+            fprintf(stderr, "[single] insert %d failed\n", i); return 1;
+        }
+    }
+    if (vindex_size(g_ix) != (size_t)NVEC) {
+        fprintf(stderr, "[single] size %zu != %d\n", vindex_size(g_ix), NVEC); return 1;
+    }
+    uint64_t ids[16]; float ds[16];
+    for (int q = 0; q < 200; q++) {
+        int k = vindex_search(g_ix, g_vecs + (size_t)((q * 7) % NVEC) * DIM, 16, 0, ids, ds);
+        if (k < 0) { fprintf(stderr, "[single] search failed at q=%d\n", q); return 1; }
+    }
+    vindex_free(g_ix); g_ix = NULL;
+    printf("[single] PASS — no memory error (this must ALWAYS pass)\n");
+    return 0;
+}
+
+/* Hazard (b) in isolation: writer + reader on a BARE index, no boundary. */
+static int run_unsynchronized(void) {
+    printf("[unsynchronized] 1 writer + 1 reader on a BARE index (TSan probe)\n");
+    printf("[unsynchronized] a race here is EXPECTED and PERMANENT — in-place HNSW\n");
+    printf("[unsynchronized] insert rewires existing elements. This is the proof that\n");
+    printf("[unsynchronized] the publication boundary must live ABOVE engram_vindex.c.\n");
+    g_ix = vindex_create(DIM, 0, 0);
+    if (!g_ix) { fprintf(stderr, "[unsynchronized] vindex_create failed\n"); return 1; }
+    for (int i = 0; i < SEED_N; i++)
+        (void)vindex_insert(g_ix, (uint64_t)i, g_vecs + (size_t)i * DIM);
+
+    pthread_t w, r;
+    if (pthread_create(&w, NULL, writer_fn, NULL) ||
+        pthread_create(&r, NULL, reader_fn, NULL)) {
+        fprintf(stderr, "[unsynchronized] pthread_create failed\n"); return 1;
+    }
+    pthread_join(w, NULL);
+    pthread_join(r, NULL);
+    vindex_free(g_ix); g_ix = NULL;
+    printf("[unsynchronized] completed — CHECK THE SANITIZER VERDICT, not this line.\n");
+    return 0;
+}
+
+/* ── hazard (a) in isolation: concurrent READS only ───────────────────────────
+ * This is what the frame-owned visited set fixes. Before that change, two
+ * vindex_search calls on one index wrote each other's epoch stamp; TSan reported
+ * the race at visited_reset and the traversal then walked bogus element indices. */
+#define NREADERS 4
+
+static int run_readers(void) {
+    printf("[readers] %d concurrent readers, NO writer, one shared index (TSan)\n", NREADERS);
+    printf("[readers] this is the visited-set regression gate — must be CLEAN.\n");
+    g_ix = vindex_create(DIM, 0, 0);
+    if (!g_ix) { fprintf(stderr, "[readers] vindex_create failed\n"); return 1; }
+    for (int i = 0; i < NVEC; i++)
+        (void)vindex_insert(g_ix, (uint64_t)i, g_vecs + (size_t)i * DIM);
+
+    pthread_t t[NREADERS];
+    for (int i = 0; i < NREADERS; i++)
+        if (pthread_create(&t[i], NULL, reader_fn, NULL)) {
+            fprintf(stderr, "[readers] pthread_create failed\n"); return 1;
+        }
+    for (int i = 0; i < NREADERS; i++) pthread_join(t[i], NULL);
+    vindex_free(g_ix); g_ix = NULL;
+    printf("[readers] completed — CHECK THE SANITIZER VERDICT, not this line.\n");
+    return 0;
+}
+
+/* ── the publication boundary, mirroring el_runtime.c ─────────────────────────
+ * Readers take the boundary SHARED and hold it across the whole search; the one
+ * owner takes it EXCLUSIVE to extend. Same shape as eg_vindex_view /
+ * eg_vindex_maintain. Note the reader's index pointer is `const VIndex*` — the
+ * compiler, not this comment, is what stops a reader inserting. */
+static pthread_rwlock_t g_pub = PTHREAD_RWLOCK_INITIALIZER;
+
+static void* pub_writer_fn(void* arg) {
+    (void)arg;
+    for (int i = SEED_N; i < NVEC; i++) {
+        pthread_rwlock_wrlock(&g_pub);
+        (void)vindex_insert(g_ix, (uint64_t)i, g_vecs + (size_t)i * DIM);
+        pthread_rwlock_unlock(&g_pub);
+    }
+    return NULL;
+}
+
+static void* pub_reader_fn(void* arg) {
+    (void)arg;
+    uint64_t ids[8]; float ds[8];
+    for (int i = 0; i < 5000; i++) {
+        pthread_rwlock_rdlock(&g_pub);
+        const VIndex* view = g_ix;                 /* immutable view */
+        (void)vindex_search(view, g_vecs + (size_t)(i % NVEC) * DIM, 8, 0, ids, ds);
+        pthread_rwlock_unlock(&g_pub);
+    }
+    return NULL;
+}
+
+static int run_published(void) {
+    printf("[published] 1 owner + %d readers through a publication boundary (TSan)\n", NREADERS);
+    printf("[published] this is the eg_vindex_view/eg_vindex_maintain gate — must be CLEAN.\n");
+    g_ix = vindex_create(DIM, 0, 0);
+    if (!g_ix) { fprintf(stderr, "[published] vindex_create failed\n"); return 1; }
+    for (int i = 0; i < SEED_N; i++)
+        (void)vindex_insert(g_ix, (uint64_t)i, g_vecs + (size_t)i * DIM);
+
+    pthread_t w, r[NREADERS];
+    if (pthread_create(&w, NULL, pub_writer_fn, NULL)) {
+        fprintf(stderr, "[published] pthread_create failed\n"); return 1;
+    }
+    for (int i = 0; i < NREADERS; i++)
+        if (pthread_create(&r[i], NULL, pub_reader_fn, NULL)) {
+            fprintf(stderr, "[published] pthread_create failed\n"); return 1;
+        }
+    pthread_join(w, NULL);
+    for (int i = 0; i < NREADERS; i++) pthread_join(r[i], NULL);
+    if (vindex_size(g_ix) != (size_t)NVEC) {
+        fprintf(stderr, "[published] size %zu != %d — the owner lost inserts\n",
+                vindex_size(g_ix), NVEC);
+        vindex_free(g_ix); g_ix = NULL; return 1;
+    }
+    vindex_free(g_ix); g_ix = NULL;
+    printf("[published] all %d inserts landed; CHECK THE SANITIZER VERDICT too.\n", NVEC);
+    return 0;
+}
+
+int main(int argc, char** argv) {
+    const char* mode = (argc > 1) ? argv[1] : "single";
+    fill_vectors();
+    int rc;
+    if      (!strcmp(mode, "single"))         rc = run_single();
+    else if (!strcmp(mode, "readers"))        rc = run_readers();
+    else if (!strcmp(mode, "unsynchronized")) rc = run_unsynchronized();
+    else if (!strcmp(mode, "published"))      rc = run_published();
+    /* back-compat: the pre-split name meant the bare writer+reader probe. */
+    else if (!strcmp(mode, "concurrent"))     rc = run_unsynchronized();
+    else {
+        fprintf(stderr, "usage: %s [single|readers|unsynchronized|published]\n", argv[0]);
+        rc = 2;
+    }
+    free(g_vecs);
+    return rc;
+}

ingest/src/ingest.el

@@ -13,7 +13,7 @@
 // relations add edges. Every node enters with PROVENANCE + grounding-level
 // + stewardship class from the moment of entry.
 //
-// transduce() is THE single mechanism — one function, polymorphic, with no
+// transduce_manifold() is THE single mechanism — one function, polymorphic, with no
 // content-type branch inside it. It does not ask whether a payload is
 // prose, structured data, or raw/opaque bytes (audio, or anything else);
 // it runs one boundary-scan-with-fixed-window-fallback chunking algorithm
@@ -401,10 +401,25 @@ fn head80(s: String) -> String {
 // truncates at the first embedded NUL, which is routine in real binary
 // bytes) is a MECHANICAL fidelity concern that belongs to whatever produced
 // `source` (see ingest_file's file_source_string below) — not a
-// content-type judgment made in here. transduce() never learns whether a
+// content-type judgment made in here. transduce_manifold() never learns whether a
 // chunk is plain text or a base64-encoded raw-byte window; every chunk is
 // handled identically either way.
-fn transduce(nodes: [String], edges: [String], source: String,
+// RENAMED transduce -> transduce_manifold (2026-08-16). Two reasons, and the
+// first is not the interesting one:
+//
+//   1. Mechanical: `transduce` is now a LANGUAGE primitive in el_runtime.h
+//      (transduce(signal, modality) -> Geometry). Every El `fn name(...)`
+//      compiles to a global C symbol with that exact name, so keeping this
+//      name here is a hard `conflicting types for 'transduce'` compile error
+//      the moment ingest.c links el_runtime.c. Measured, not anticipated.
+//
+//   2. Actual: this function was never signal->geometry. It chunks already-
+//      extracted content and PACKS it into a node+edge manifold — a real
+//      operation, but one layer up, and it had taken the name that belongs to
+//      the primitive underneath it. `transduce` is where a signal becomes
+//      geometry; `transduce_manifold` is where extracted content becomes
+//      structure. Nothing about this function's behaviour changed.
+fn transduce_manifold(nodes: [String], edges: [String], source: String,
              prov: String, ground: String, steward: String,
              root_lid: String, root_title: String) -> [String] {
     let tagbase: String = "prov:" + prov + " ground:" + ground + " steward:" + steward
@@ -531,8 +546,8 @@ fn default_steward() -> String {
 // trustworthy verbatim. When they don't (silent truncation happened),
 // rebuild the payload as base64-encoded fixed-size windows read directly
 // off disk (fs_read_b64_chunk — binary-safe in C), joined with the same
-// "\n\n" boundary marker transduce()'s generic scan already looks for, so
-// transduce() sees one ordinary boundary-delimited payload and runs its one
+// "\n\n" boundary marker transduce_manifold()'s generic scan already looks for, so
+// transduce_manifold() sees one ordinary boundary-delimited payload and runs its one
 // algorithm on it exactly as it would on prose — it never learns that a
 // fidelity problem occurred upstream, let alone why.
 fn file_source_string(path: String, text: String, real_size: Int) -> String {
@@ -541,7 +556,7 @@ fn file_source_string(path: String, text: String, real_size: Int) -> String {
     // 3072 raw bytes -> 4096 base64 chars (3 divides evenly into base64's
     // 3-byte/4-char ratio); keeps each resulting node's content a clean,
     // bounded, low-kilobytes unit, same order of magnitude as the fixed
-    // fallback window in transduce() itself.
+    // fallback window in transduce_manifold() itself.
     let win: Int = 3072
     let out: String = ""
     let off: Int = 0
@@ -561,7 +576,7 @@ fn file_source_string(path: String, text: String, real_size: Int) -> String {
 }
 
 // ingest one file -> report JSON. Uniform for every file regardless of
-// extension or content — transduce() decides nothing about content-type, so
+// extension or content — transduce_manifold() decides nothing about content-type, so
 // neither does this function; it only decides whether the raw bytes made it
 // through the read intact (file_source_string), which is a fidelity
 // question, not a format one.
@@ -573,14 +588,14 @@ fn ingest_file(path: String) -> String {
         return "{\"error\":\"empty or unreadable\",\"path\":" + j_q(path) + "}"
     }
     let prov: String = "file:" + path
-    let packed: [String] = transduce(el_list_empty(), el_list_empty(),
+    let packed: [String] = transduce_manifold(el_list_empty(), el_list_empty(),
         source, prov, default_ground(), default_steward(),
         "doc:" + basename(path), basename(path))
     return merge_packed(packed)
 }
 
 // ingest a directory: walk one level, ingest every file found, aggregate.
-// No extension filter — transduce() handles any payload uniformly now, so
+// No extension filter — transduce_manifold() handles any payload uniformly now, so
 // there is no content-type gate at the directory boundary either.
 fn ingest_dir(path: String) -> String {
     let entries: [String] = fs_list(path)
@@ -615,7 +630,7 @@ fn ingest_dir(path: String) -> String {
 fn ingest_url(url: String) -> String {
     let body: String = http_get(url)
     if str_eq(body, "") { return "{\"error\":\"empty fetch\",\"url\":" + j_q(url) + "}" }
-    let packed: [String] = transduce(el_list_empty(), el_list_empty(),
+    let packed: [String] = transduce_manifold(el_list_empty(), el_list_empty(),
         body, "url:" + url, "extracted", "public-web",
         "url:" + url, url)
     return merge_packed(packed)
@@ -630,7 +645,7 @@ fn ingest_llm(query: String) -> String {
     let resp: String = http_post_json("http://127.0.0.1:11434/api/generate", body)
     let answer: String = json_get_string(resp, "response")
     if str_eq(answer, "") { return "{\"error\":\"no model response\"}" }
-    let packed: [String] = transduce(el_list_empty(), el_list_empty(),
+    let packed: [String] = transduce_manifold(el_list_empty(), el_list_empty(),
         answer, "llm:" + model + ":" + query, "candidate-provisional", "guide-provisional",
         "llm:" + query, "guide answer: " + query)
     return merge_packed(packed)
@@ -682,7 +697,7 @@ fn ingest_stream(path: String) -> String {
 // It is NOT a content-type flag: it says nothing about what's inside the
 // bytes once fetched, and none of the five ingest_* functions it selects
 // among interpret their payload differently by content shape anymore —
-// they all hand off to the single, format-agnostic transduce(). The old
+// they all hand off to the single, format-agnostic transduce_manifold(). The old
 // "structured" value (a caller-declared alias for "file", used only to hint
 // the now-removed JSON-vs-prose branch) is gone along with that branch.
 let kind: String = env("INGEST_KIND")

lang/AGENTS.md

@@ -62,34 +62,45 @@ This is where almost all work belongs. El programs are source files that get com
 
 This is the self-contained C OS-boundary layer. It provides the `__`-prefixed primitives that compiled El programs call: libcurl HTTP, pthreads, filesystem I/O, arena allocation, etc. It is **not generated** — it is maintained by hand.
 
-The old `el_runtime.c` has been archived to `runtime/legacy/`. The runtime is now native El (`runtime/*.el`). `el_seed.c` replaces `el_runtime.c` as the sole C compilation dependency.
+The runtime is native El (`runtime/*.el`) over a C OS-boundary. **Status (verified 2026-08-15):** the migration to a seed-only boundary is *in progress, not done*. Two files exist:
+- `runtime/el_runtime.c` (~860 KB) — **LIVE**. Holds the engram store (`EngramStore engram_global`) plus the `http_*`/`json_*`/`state_*`/`engram_*` impls. It is the authoritative single-file link target for the compiler, and `tools/install.sh` compiles it into `libel.a`. This is where a new C builtin's *implementation* must currently live to be linkable.
+- `runtime/el_seed.c` — the intended hand-maintained `__`-prefixed seed (thin wrappers over the above). It is compiled alongside `el_runtime.c` by `tools/install.sh`, but does **not** compile standalone yet (see the build-path caveat under "Rebuilding the Compiler").
 
-**Only edit `el_seed.c` when you genuinely need OS-level access** (raw sockets, GPU calls, new libcurl features). For everything else, write El.
+**Only edit these when you genuinely need OS-level access** (raw sockets, GPU calls, new libcurl features, a new engram store op). For everything else, write El.
 
-When you do add a C builtin:
-1. Add the C function to `el_seed.c`
-2. Declare it in `el_seed.h`
-3. Add it to the `builtin_arity` table in `el-compiler/src/codegen.el` (so the compiler knows the arg count)
-4. Rebuild the elc binary (see below)
+When you add a C builtin (verbatim-emit recipe — the El name is emitted as the exact C symbol; `builtin_arity` is an arity guard only, not a dispatch table):
+1. Implement the C function in `el_runtime.c` (and declare it in `el_runtime.h`).
+2. Add a `__`-prefixed thin wrapper in `el_seed.c` and declare it in `el_seed.h`.
+3. Add the name to `builtin_arity` in `el-compiler/src/codegen.el` — add **both** the plain and `__`-prefixed spellings.
+4. Rebuild the elc binary (see below) and confirm the self-host fixpoint is byte-identical.
+
+Worked example: the `engram_assert_json` (op_assert seam) and `engram_node_full_in`/`engram_connect_in` (purview write-side) primitives added 2026-08-15 follow exactly this recipe.
 
 ---
 
 ## Rebuilding the Compiler
 
-After changing any `.el` source in `el-compiler/src/`:
+After changing any `.el` source in `el-compiler/src/` (run from the `lang/` dir):
 
 ```bash
-cd /Users/will/Development/neuron-technologies/foundation/el
+# 1. Stage2: current elc compiles the (modified) compiler to C
 ./dist/platform/elc elc-cli.el > elc-new.c
+# 2. Build the new compiler. The C link target is el_runtime.c — it holds the
+#    engram store + http/json/state impls the compiler output calls. el_runtime.c
+#    self-hosts elc on its own; el_seed.c is the (aspirational) seed layer and does
+#    NOT compile standalone under clang (missing prototypes for the el_runtime.c
+#    symbols it wraps — see caveat below), so link el_runtime.c here.
 cc -std=c11 -I runtime -lcurl -lpthread \
    -o dist/platform/elc-new \
-   elc-new.c runtime/el_seed.c
-# Verify self-hosting:
+   elc-new.c runtime/el_runtime.c
+# 3. Verify self-hosting FIXPOINT (stage3 == stage2 output, byte-identical):
 ./dist/platform/elc-new elc-cli.el > elc-verify.c
-diff elc-new.c elc-verify.c   # should be identical
+diff elc-new.c elc-verify.c   # must be identical
 mv dist/platform/elc-new dist/platform/elc
 ```
 
+> **Build-path caveat (verified 2026-08-15).** `el_seed.c` is the intended hand-maintained OS-boundary seed, but it does **not** compile standalone under modern clang: it wraps ~16 unprefixed `el_runtime.c` symbols (`http_serve`, `json_*`, `state_*`, `http_response`) without prototypes, and clang treats implicit declarations as errors (C99+). The productionised install (`tools/install.sh`) builds `libel.a` from **both** `el_seed.o` + `el_runtime.o` together, which is why linking succeeds there. To make `el_seed.c` build on its own, add prototypes for those symbols (or `#include "el_runtime.h"`, reconciling the `__http_serve` return-type mismatch first). Until then, `el_runtime.c` is the authoritative single-file link target for the compiler.
+
 After changing `el_seed.c` only (no El source changes), rebuild downstream programs but do NOT need to rebuild the compiler binary itself — the seed is linked at the application level, not the compiler level.
 
 ---

lang/el-compiler/src/codegen.el

@@ -862,10 +862,23 @@ fn cg_expr(expr: Map<String, Any>) -> String {
             // arithmetic BinOp (or vice-versa). Without this check the
             // fallthrough to str_eq produces str_eq(int_value, int_value)
             // which reads the integer as a char* and segfaults.
+            // EITHER side provably Int is enough. Requiring BOTH meant a call
+            // whose return type codegen cannot infer poisoned the operator:
+            //     getint(5) == a      ->  str_eq(getint(5), a)
+            // even with `a` declared Int. str_eq then reads an integer as a
+            // char* and segfaults. Only an integer LITERAL on one side forced
+            // the numeric form, so the bug was invisible in the common case.
+            //
+            // Loosening to OR is strictly safer: when one side is a known Int,
+            // str_eq is always wrong (it dereferences that int), while numeric
+            // comparison is at worst a wrong answer on an already ill-typed
+            // program. When neither side is Int nothing changes, so string
+            // comparison is untouched.
             if is_int_expr(left) {
-                if is_int_expr(right) {
-                    return "(" + left_c + " == " + right_c + ")"
-                }
+                return "(" + left_c + " == " + right_c + ")"
+            }
+            if is_int_expr(right) {
+                return "(" + left_c + " == " + right_c + ")"
             }
             // Float literal or negative float literal: use plain == (bit-equal
             // el_val_t comparison). This handles `r0 == 3.0`, `neg == -3.0`, etc.
@@ -921,10 +934,12 @@ fn cg_expr(expr: Map<String, Any>) -> String {
             }
             // Same mixed Ident/BinOp fix as EqEq: use is_int_expr to detect
             // integer-typed operands before falling through to !str_eq.
+            // Either side Int is enough — see the EqEq note above.
             if is_int_expr(left) {
-                if is_int_expr(right) {
-                    return "(" + left_c + " != " + right_c + ")"
-                }
+                return "(" + left_c + " != " + right_c + ")"
+            }
+            if is_int_expr(right) {
+                return "(" + left_c + " != " + right_c + ")"
             }
             // Float-typed operands use plain != (bit-equal comparison).
             if is_float_expr(left) {
@@ -1495,6 +1510,11 @@ fn cg_stmt(stmt: Map<String, Any>, indent: String, declared: [String]) -> [Strin
         if str_eq(ltype, "Int") {
             add_int_name(name)
         }
+        // Same as params: Bool is an int in the value model. Without this a
+        // `let ok: Bool = ...` compared to another Bool lowered to str_eq.
+        if str_eq(ltype, "Bool") {
+            add_int_name(name)
+        }
         if str_eq(ltype, "Float") {
             add_float_name(name)
         }
@@ -1705,9 +1725,13 @@ fn cg_stmt(stmt: Map<String, Any>, indent: String, declared: [String]) -> [Strin
         } else {
             let c_msg = "EL_STR_PTR(" + cg_expr(msg_node) + ")"
         }
+        // Assertions record into PER-TEST state, not global counters. The test
+        // is the unit of result; a global pass/fail tally cannot say which test
+        // failed or whether a test ran at all. Reporting is the runner's job —
+        // nothing is printed here.
         emit_line(indent + "if (!(" + c_cond + ")) {")
-        emit_line(indent + "  __el_test_fail(__el_cur_test, " + c_msg + "); __el_fail++;")
-        emit_line(indent + "} else { __el_pass++; }")
+        emit_line(indent + "  __el_test_fail(" + c_msg + ");")
+        emit_line(indent + "} else { __el_cur_asserts++; }")
         return declared
     }
 
@@ -2602,6 +2626,17 @@ fn builtin_arity(name: String) -> Int {
     // LSP seed primitives
     if str_eq(name, "__read_n") { return 1 }
     if str_eq(name, "__print_raw") { return 1 }
+    // Test-registry accessors. These are not runtime builtins — they are
+    // GENERATED into the same translation unit by the --test path below, one
+    // set per test binary. They are declared here so the El-side runner in
+    // runtime/eltest.el can call them with a known arity.
+    if str_eq(name, "__el_reg_count") { return 0 }
+    if str_eq(name, "__el_reg_name") { return 1 }
+    if str_eq(name, "__el_reg_invoke") { return 1 }
+    if str_eq(name, "__el_reg_last_ns") { return 0 }
+    if str_eq(name, "__el_reg_msg") { return 0 }
+    if str_eq(name, "__el_reg_asserts") { return 0 }
+    if str_eq(name, "__el_opt_json") { return 0 }
     // String
     if str_eq(name, "el_str_concat") { return 2 }
     if str_eq(name, "str_eq") { return 2 }
@@ -2760,7 +2795,15 @@ fn builtin_arity(name: String) -> Int {
     if str_eq(name, "__engram_neighbors_filtered") { return 3 }
     if str_eq(name, "__engram_activate") { return 2 }
     if str_eq(name, "__engram_activate_json") { return 2 }
+    if str_eq(name, "__engram_op_assert_json") { return 2 }
+    if str_eq(name, "__engram_node_full_in") { return 9 }
+    if str_eq(name, "__engram_connect_in") { return 5 }
     if str_eq(name, "__engram_scan_nodes_json") { return 2 }
+    if str_eq(name, "__engram_edges_json") { return 2 }
+    if str_eq(name, "__engram_pool_stats_json") { return 0 }
+    if str_eq(name, "__el_alloc_count") { return 0 }
+    if str_eq(name, "__el_alloc_bytes") { return 0 }
+    if str_eq(name, "__el_peak_rss") { return 0 }
     if str_eq(name, "__generate") { return 1 }
     // Filesystem
     if str_eq(name, "fs_read") { return 1 }
@@ -2859,9 +2902,18 @@ fn builtin_arity(name: String) -> Int {
     if str_eq(name, "engram_get_node_by_label") { return 1 }
     if str_eq(name, "engram_search_json") { return 2 }
     if str_eq(name, "engram_scan_nodes_json") { return 2 }
+    if str_eq(name, "engram_edges_json") { return 2 }
+    if str_eq(name, "engram_pool_stats_json") { return 0 }
+    if str_eq(name, "el_alloc_count") { return 0 }
+    if str_eq(name, "el_alloc_bytes") { return 0 }
+    if str_eq(name, "el_peak_rss") { return 0 }
+    if str_eq(name, "el_black_box") { return 1 }
     if str_eq(name, "engram_neighbors_json") { return 3 }
     if str_eq(name, "engram_activate_json") { return 2 }
     if str_eq(name, "engram_stats_json") { return 0 }
+    if str_eq(name, "engram_op_assert_json") { return 2 }
+    if str_eq(name, "engram_node_full_in") { return 9 }
+    if str_eq(name, "engram_connect_in") { return 5 }
     // LLM
     if str_eq(name, "llm_call") { return 2 }
     if str_eq(name, "llm_call_system") { return 3 }
@@ -3081,6 +3133,15 @@ fn build_int_names_for_params(params: [Map<String, Any>]) -> Bool {
         if str_eq(ptype, "Int") {
             add_int_name(pname)
         }
+        // Bool is an integer in the value model (type_to_c maps Bool -> "int";
+        // el_runtime.h: "Bool -> el_val_t (0 = false, nonzero = true)"), but
+        // Bool names were registered nowhere. So `cond == want` between two
+        // Bool params fell through to str_eq and dereferenced 0 or 1 as a
+        // char* — an immediate segfault. Track them as int-like, which is what
+        // they are.
+        if str_eq(ptype, "Bool") {
+            add_int_name(pname)
+        }
         if str_eq(ptype, "Float") {
             add_float_name(pname)
         }
@@ -4100,13 +4161,36 @@ fn codegen_streaming(tokens: [Any], sigs: [Map<String, Any>], source: String) ->
     // Emit test harness preamble (counters, fail printer) when in test mode.
     if test_is_mode {
         emit_line("#include <stdio.h>")
+        emit_line("#include <string.h>")
+        emit_line("#include <time.h>")
         emit_blank()
-        emit_line("static int __el_pass = 0, __el_fail = 0;")
+        // Per-test result state. Reset by __el_reg_invoke before each test, so
+        // every test gets its own record rather than contributing to a global
+        // tally. The first failure message is retained; later ones only bump
+        // the count, which keeps the common case allocation-free.
+        emit_line("static int __el_cur_fails = 0;")
+        emit_line("static int __el_cur_asserts = 0;")
+        emit_line("static char __el_cur_msg[512] = \"\";")
         emit_line("static const char *__el_cur_test = \"(none)\";")
-        emit_line("static void __el_test_fail(const char *test, const char *msg) {")
-        emit_line("  fprintf(stderr, \"FAIL  %-40s %s\\n\", test, msg);")
+        emit_line("static void __el_test_fail(const char *msg) {")
+        emit_line("  if (__el_cur_fails == 0 && msg) {")
+        emit_line("    snprintf(__el_cur_msg, sizeof __el_cur_msg, \"%s\", msg);")
+        emit_line("  }")
+        emit_line("  __el_cur_fails++; __el_cur_asserts++;")
         emit_line("}")
         emit_blank()
+        // Forward declarations for the registry accessors. The definitions are
+        // emitted at the END of the unit (they reference the test functions,
+        // which do not exist yet at this point), but the El-side runner is
+        // compiled in between and calls them — so it needs the prototypes here.
+        emit_line("el_val_t __el_reg_count(void);")
+        emit_line("el_val_t __el_reg_name(el_val_t i);")
+        emit_line("el_val_t __el_reg_invoke(el_val_t i);")
+        emit_line("el_val_t __el_reg_last_ns(void);")
+        emit_line("el_val_t __el_reg_msg(void);")
+        emit_line("el_val_t __el_reg_asserts(void);")
+        emit_line("el_val_t __el_opt_json(void);")
+        emit_blank()
     }
 
     // Streaming parse-emit loop.
@@ -4302,17 +4386,72 @@ fn codegen_streaming(tokens: [Any], sigs: [Map<String, Any>], source: String) ->
         el_release(sigs)
 
         let test_arena_mark: Any = el_arena_push()
+        let tn: Int = native_list_len(test_c_names)
+
+        // ── Generated test registry ──────────────────────────────────────────
+        // Discovery happens HERE, at compile time. The runner never searches
+        // for tests; it walks this table. That ordering — discovery strictly
+        // before execution — is what makes --list, filtering, sharding and
+        // per-test reporting possible later, and it is why the old harness
+        // (which inlined direct calls into main) could not have any of them.
+        emit_line("typedef void (*__el_test_fp)(void);")
+        emit_line("typedef struct { const char *name; __el_test_fp fn; } __el_test_entry;")
+        emit_line("static const __el_test_entry __el_registry[] = {")
+        let ri: Int = 0
+        while ri < tn {
+            let r_name: String = native_list_get(test_names, ri)
+            let r_cfn: String = native_list_get(test_c_names, ri)
+            emit_line("  { \"" + c_escape(r_name) + "\", " + r_cfn + " },")
+            let ri = ri + 1
+        }
+        // Trailing sentinel keeps the array non-empty when a file declares no
+        // tests (a zero-length array is not valid C).
+        emit_line("  { 0, 0 }")
+        emit_line("};")
+        emit_line("static const int __el_registry_n = " + int_to_str(tn) + ";")
+        emit_blank()
+        emit_line("static long long __el_last_ns = 0;")
+        emit_line("static int __el_opt_json_v = 0;")
+        emit_blank()
+
+        // ── Index-based accessors ────────────────────────────────────────────
+        // El has no function pointers, so the runner works purely in indices.
+        // This is the whole seam between generated C and the El-side runner.
+        emit_line("el_val_t __el_reg_count(void) { return (el_val_t)(int64_t)__el_registry_n; }")
+        emit_line("el_val_t __el_reg_name(el_val_t i) {")
+        emit_line("  int64_t k = (int64_t)i;")
+        emit_line("  if (k < 0 || k >= __el_registry_n) return EL_STR(\"\");")
+        emit_line("  return EL_STR(__el_registry[k].name);")
+        emit_line("}")
+        // Timing is taken immediately around the call, in C, on the MONOTONIC
+        // clock — never the wall clock, which can step backwards under NTP.
+        emit_line("el_val_t __el_reg_invoke(el_val_t i) {")
+        emit_line("  int64_t k = (int64_t)i;")
+        emit_line("  if (k < 0 || k >= __el_registry_n) return 0;")
+        emit_line("  __el_cur_fails = 0; __el_cur_asserts = 0; __el_cur_msg[0] = '\\0';")
+        emit_line("  __el_cur_test = __el_registry[k].name;")
+        emit_line("  struct timespec _t0, _t1;")
+        emit_line("  clock_gettime(CLOCK_MONOTONIC, &_t0);")
+        emit_line("  __el_registry[k].fn();")
+        emit_line("  clock_gettime(CLOCK_MONOTONIC, &_t1);")
+        emit_line("  __el_last_ns = (long long)(_t1.tv_sec - _t0.tv_sec) * 1000000000LL")
+        emit_line("              + (long long)(_t1.tv_nsec - _t0.tv_nsec);")
+        emit_line("  return (el_val_t)(int64_t)__el_cur_fails;")
+        emit_line("}")
+        emit_line("el_val_t __el_reg_last_ns(void) { return (el_val_t)(int64_t)__el_last_ns; }")
+        emit_line("el_val_t __el_reg_msg(void) { return EL_STR(__el_cur_msg); }")
+        emit_line("el_val_t __el_reg_asserts(void) { return (el_val_t)(int64_t)__el_cur_asserts; }")
+        emit_line("el_val_t __el_opt_json(void) { return (el_val_t)(int64_t)__el_opt_json_v; }")
+        emit_blank()
+
+        // main() delegates to the El-side runner. Everything above this line is
+        // generated glue; all reporting logic lives in runtime/eltest.el.
         emit_line("int main(int _argc, char **_argv) {")
         emit_line("  el_runtime_init_args(_argc, _argv);")
-        let ti: Int = 0
-        let tn: Int = native_list_len(test_c_names)
-        while ti < tn {
-            let tc_name: String = native_list_get(test_c_names, ti)
-            emit_line("  " + tc_name + "();")
-            let ti = ti + 1
-        }
-        emit_line("  printf(\"%d passed, %d failed\\n\", __el_pass, __el_fail);")
-        emit_line("  return __el_fail;")
+        emit_line("  for (int _i = 1; _i < _argc; _i++) {")
+        emit_line("    if (strcmp(_argv[_i], \"--json\") == 0) __el_opt_json_v = 1;")
+        emit_line("  }")
+        emit_line("  return (int)(int64_t)el_test_main();")
         emit_line("}")
         el_arena_pop(test_arena_mark)
         el_release(test_names)

lang/el-compiler/src/compiler.el

@@ -419,6 +419,22 @@ fn resolve_imports(src_path: String) -> String {
     if !str_eq(already, "") { return "" }
     state_set(seen_key, "1")
 
+    // A missing file must be a hard error, never an empty string.
+    //
+    // fs_read returns "" both for "file is empty" and "file does not exist", and
+    // this function used the value without distinguishing them. So a broken
+    // import path — a typo, a moved file, a relative path resolved from the
+    // wrong working directory — compiled CLEANLY: exit 0, empty stderr, and a
+    // program silently missing everything it imported. Observed 2026-08-15:
+    // eleven consecutive "successful" compiles that had included no runtime at
+    // all, and a wrong conclusion drawn from them before anyone noticed.
+    //
+    // Missing dependency, confident success. fs_exists separates the two cases,
+    // so a genuinely empty file still resolves to "" and is fine.
+    if !fs_exists(src_path) {
+        println("elc: cannot resolve import: " + src_path)
+        exit_program(1)
+    }
     let source: String = fs_read(src_path)
     let dir: String = dirname_of(src_path)
     let lines: [String] = str_split(source, "\n")

lang/examples/transduce.el

@@ -0,0 +1,213 @@
+// transduce.el — geometry as a first-class El value, and a realizer written
+// in El. Runnable: this is the worked example for the transduce surface, and
+// it doubles as an executable proof because it checks every claim it makes.
+//
+//   elc lang/examples/transduce.el > transduce.c
+//   cc -std=c11 -O2 -I lang/runtime -o transduce transduce.c \
+//      lang/runtime/el_runtime.c lang/runtime/el_seed.c \
+//      lang/runtime/engram_*.c -lcurl -lpthread -lm
+//   ./transduce          # exits 0 only if every check passes
+//
+// (A `test "..."` form of the same checks lives in
+// lang/tests/native/test_transduce.el, for when the native harness is
+// repaired — the shipped elc currently emits calls to __el_reg_count and
+// friends without emitting their definitions, which breaks every native test
+// equally, test_math.el included. Verified 2026-08-16, unrelated to this work.)
+//
+// WHY THIS EXISTS. Until 2026-08-16 no El ingest path could carry a vector:
+// nodes took text, and geometry was DERIVED from that text. Text was the
+// mandatory entry medium, so any non-text modality had to be DESCRIBED in
+// prose first and the geometry we reasoned over was the geometry OF THE
+// DESCRIPTION, not of the signal. Two things fix that, and both are shown
+// below: geometry is a VALUE that carries its own width, and a REALIZER is an
+// ordinary El function — so admitting a new modality never requires a runtime
+// patch.
+//
+// COMPARISON DISCIPLINE (measured, not stylistic): elc lowers `a == b`
+// numerically only when both operand NAMES are in the per-function int-name
+// set that `let x: Int` populates. A bare `f(x) == 0` is not a registered
+// name and lowers to str_eq — strcmp on two integers as pointers. `<` and `>`
+// lower directly with no inference, so truthiness is written `> 0` / `< 1`.
+
+// ── A realizer, written entirely in El ──────────────────────────────────────
+// Not in the runtime. Not known to the compiler. Registered by NAME and
+// dispatched to through transduce(). That is the whole claim.
+fn tone_realizer(signal: String) -> Geometry {
+    let g: Geometry = geometry_new(4)
+    let n: Int = str_len(signal)
+    let a: Int = geometry_set(g, 0, int_to_float(n))
+    let b: Int = geometry_set(g, 1, int_to_float(n * 2))
+    let c: Int = geometry_set(g, 2, int_to_float(n * 3))
+    let d: Int = geometry_set(g, 3, int_to_float(n * 4))
+    g
+}
+
+// A second modality, to show the registry keys on modality rather than just
+// returning whatever was registered last.
+fn pulse_realizer(signal: String) -> Geometry {
+    let g: Geometry = geometry_new(2)
+    let a: Int = geometry_set(g, 0, 1.0)
+    let b: Int = geometry_set(g, 1, 0.0)
+    g
+}
+
+// A deliberately BROKEN realizer: returns something that is not a Geometry.
+fn bogus_realizer(signal: String) -> Geometry {
+    return 12345
+}
+
+// Fails FAST rather than accumulating a count, for a measured reason: a first
+// cut wrote `let fails: Int = fails + check(...)` and `+` lowered to STRING
+// CONCAT, because elc dispatches `+` on whether both operands are known-Int and
+// a user-defined fn call is not — so the counter printed 4343632752, a pointer.
+// Nothing was wrong with the checks; the tally was lying. Exiting at the first
+// failure needs no arithmetic at all, so there is nothing left to get wrong.
+fn check(ok: Int, label: String) -> Int {
+    if ok > 0 {
+        println("  ok    " + label)
+        return 0
+    }
+    println("  FAIL  " + label)
+    exit(1)
+    return 1
+}
+
+fn near(a: Float, b: Float) -> Int {
+    let d: Float = a - b
+    if d > 0.001 { return 0 }
+    if d < -0.001 { return 0 }
+    return 1
+}
+
+fn eq_int(a: Int, b: Int) -> Int {
+    if a == b { return 1 }
+    return 0
+}
+
+fn main() -> Void {
+    println("geometry is a value that carries its own width")
+    let g8: Geometry = geometry_new(8)
+    let _c: Int = check(geometry_is(g8), "geometry_new returns a live Geometry")
+    let d8: Int = geometry_dim(g8)
+    let _c: Int = check(eq_int(d8, 8), "a Geometry carries its own width (8)")
+    let _c: Int = check(geometry_free(g8), "geometry_free reports what it did")
+
+    println("nonsense is refused — with no arbitrary max-dim bound")
+    // #141 needed `dim <= 8192` only to bound an allocation sized from a
+    // caller's CLAIM about a string's length. A value that carries its own
+    // width has nothing left to validate.
+    let z: Geometry = geometry_new(0)
+    let zi: Int = geometry_is(z)
+    let _c: Int = check(1 - zi, "dim 0 is not a geometry")
+    let ng: Geometry = geometry_new(-4)
+    let ngi: Int = geometry_is(ng)
+    let _c: Int = check(1 - ngi, "negative dim is not a geometry")
+    let nd: Int = geometry_dim(0)
+    let _c: Int = check(1 - nd, "geometry_dim of a non-geometry is 0, not a crash")
+    let nf: Int = geometry_free(0)
+    let _c: Int = check(1 - nf, "geometry_free of a non-geometry is a no-op")
+
+    println("components round-trip, and out-of-range is refused")
+    let g3: Geometry = geometry_new(3)
+    let s0: Int = geometry_set(g3, 0, 1.5)
+    let s1: Int = geometry_set(g3, 1, -2.5)
+    let _c: Int = check(s0, "set in range succeeds")
+    let oob: Int = geometry_set(g3, 3, 9.0)
+    let _c: Int = check(1 - oob, "set out of range is refused, not silently dropped")
+    let _c: Int = check(near(geometry_get(g3, 0), 1.5), "component 0 round-trips")
+    let _c: Int = check(near(geometry_get(g3, 1), -2.5), "component 1 round-trips (negative)")
+    let ff3: Int = geometry_free(g3)
+
+    println("hex is an EDGE adapter, and derives its own width")
+    // little-endian float32: 1.0 = 0000803f, 2.0 = 00000040
+    let gh: Geometry = geometry_from_f32le_hex("0000803f00000040")
+    let _c: Int = check(geometry_is(gh), "valid hex decodes to a Geometry")
+    let dh: Int = geometry_dim(gh)
+    let _c: Int = check(eq_int(dh, 2), "width DERIVED from input, never supplied")
+    let _c: Int = check(near(geometry_get(gh, 0), 1.0), "first component decoded")
+    let _c: Int = check(near(geometry_get(gh, 1), 2.0), "second component decoded")
+    let back: String = geometry_to_f32le_hex(gh)
+    let _c: Int = check(str_eq(back, "0000803f00000040"), "hex round-trips exactly")
+    let ffh: Int = geometry_free(gh)
+
+    println("malformed hex is refused")
+    let he: Geometry = geometry_from_f32le_hex("")
+    let hei: Int = geometry_is(he)
+    let _c: Int = check(1 - hei, "empty hex is not a geometry")
+    let hr: Geometry = geometry_from_f32le_hex("0000803f0000")
+    let hri: Int = geometry_is(hr)
+    let _c: Int = check(1 - hri, "length not a multiple of 8 is refused")
+    let hn: Geometry = geometry_from_f32le_hex("zzzzzzzz")
+    let hni: Int = geometry_is(hn)
+    let _c: Int = check(1 - hni, "non-hex characters are refused")
+
+    println("a realizer declared in El is a first-class realizer")
+    let reg: Int = realizer_register("tone", "tone_realizer")
+    let _c: Int = check(reg, "an El fn registers as a realizer BY NAME")
+    let _c: Int = check(realizer_has("tone"), "the modality now has an organ")
+    let gt: Geometry = transduce("aaa", "tone")
+    let _c: Int = check(geometry_is(gt), "transduce returns real geometry")
+    let dt: Int = geometry_dim(gt)
+    let _c: Int = check(eq_int(dt, 4), "the El realizer determined the width, not the runtime")
+    // str_len("aaa") == 3, so component 0 must be 3.0 — proof the signal
+    // actually reached the El function rather than a stub answering for it.
+    let _c: Int = check(near(geometry_get(gt, 0), 3.0), "the signal REACHED the El realizer")
+    let fft: Int = geometry_free(gt)
+
+    println("distinct signals transduce to distinct geometry")
+    let g1: Geometry = transduce("aa", "tone")
+    let g2: Geometry = transduce("aaaaa", "tone")
+    let a1: Float = geometry_get(g1, 0)
+    let a2: Float = geometry_get(g2, 0)
+    // 5 - 2 = 3. If transduction were a stub these would be equal.
+    let _c: Int = check(near(a2 - a1, 3.0), "different signals produce different geometry")
+    let ff1: Int = geometry_free(g1)
+    let ff2: Int = geometry_free(g2)
+
+    println("the registry keys on modality")
+    let r2: Int = realizer_register("pulse", "pulse_realizer")
+    let _c: Int = check(r2, "a second modality registers independently")
+    let mt: Geometry = transduce("aaa", "tone")
+    let mp: Geometry = transduce("aaa", "pulse")
+    let mdt: Int = geometry_dim(mt)
+    let mdp: Int = geometry_dim(mp)
+    let _c: Int = check(eq_int(mdt, 4), "tone still routes to its own realizer")
+    let _c: Int = check(eq_int(mdp, 2), "pulse routes to a different realizer")
+    let ffm1: Int = geometry_free(mt)
+    let ffm2: Int = geometry_free(mp)
+
+    println("no organ is reported as no organ")
+    // A modality with no realizer must transduce to NOTHING. It must never
+    // fall back to embedding a description of the signal and calling that
+    // perception — that silent substitution is the defect this all exists to end.
+    let eh: Int = realizer_has("echolocation")
+    let _c: Int = check(1 - eh, "unregistered modality has no organ")
+    let ge: Geometry = transduce("anything", "echolocation")
+    let gei: Int = geometry_is(ge)
+    let _c: Int = check(1 - gei, "no realizer means NO geometry, not fake geometry")
+
+    println("an unresolvable realizer name fails at WIRING time")
+    let bad: Int = realizer_register("ghost", "no_such_function_anywhere")
+    let _c: Int = check(1 - bad, "unresolvable realizer name is a registration failure")
+    let gh2: Int = realizer_has("ghost")
+    let _c: Int = check(1 - gh2, "and nothing gets registered")
+
+    println("a realizer returning non-geometry transduces nothing")
+    let rb: Int = realizer_register("bogus", "bogus_realizer")
+    let _c: Int = check(rb, "the symbol resolves, so registration succeeds")
+    let gb: Geometry = transduce("x", "bogus")
+    let gbi: Int = geometry_is(gb)
+    let _c: Int = check(1 - gbi, "contract enforced at the boundary: nothing handed back")
+
+    println("norm lets a caller check a realizer emitted signal, not zeros")
+    let gn: Geometry = geometry_new(2)
+    let _c: Int = check(near(geometry_norm(gn), 0.0), "a fresh geometry is zero — norm says so")
+    let n0: Int = geometry_set(gn, 0, 3.0)
+    let n1: Int = geometry_set(gn, 1, 4.0)
+    let _c: Int = check(near(geometry_norm(gn), 5.0), "3-4-5: norm is 5")
+    let ffn: Int = geometry_free(gn)
+
+    // Reaching here means nothing called exit(1) along the way.
+    println("")
+    println("all checks passed")
+}

lang/runtime/build_vindex_bench.sh

@@ -0,0 +1,51 @@
+#!/bin/bash
+# build_vindex_bench.sh — build the vindex_bench oracle/proof harness, with
+# the real ggml + hand-rolled-Metal batch-cosine strategies on Darwin and a
+# zero-dependency CPU-only stub everywhere else. Mirrors the two-step recipe
+# documented in vindex_bench.c's own header comment; this script exists so
+# that recipe is one command, not a copy-pasted paragraph.
+#
+# Darwin build links FOUR strategy translation units:
+#   eg_cosine_batch.c                    — the Factory (always)
+#   eg_cosine_batch_strategy_cpu.c       — universal fallback (always)
+#   eg_cosine_batch_strategy_ggml.c      — ggml + dynamic Metal backend plugin
+#   eg_cosine_batch_strategy_metal_hand.m — PR #114's original hand-rolled
+#                                            Metal shader, preserved as one
+#                                            selectable strategy
+# plus -DEG_HAVE_STRATEGY_GGML -DEG_HAVE_STRATEGY_METAL_HAND so the Factory
+# (and vindex_bench.c's own direct strategy comparison) knows both exist.
+#
+# ggml is resolved via `brew --prefix ggml` when available (portable across
+# Intel /usr/local and Apple Silicon /opt/homebrew installs), falling back to
+# /opt/homebrew if brew isn't on PATH. Override with GGML_PREFIX=... env var.
+#
+# Usage: ./build_vindex_bench.sh [output_path]
+set -euo pipefail
+cd "$(dirname "$0")"
+OUT="${1:-./vindex_bench}"
+CC="${CC:-cc}"
+
+if [ "$(uname -s)" = "Darwin" ]; then
+  GGML_PREFIX="${GGML_PREFIX:-$(brew --prefix ggml 2>/dev/null || echo /opt/homebrew)}"
+  echo "== Darwin: building with the ggml + hand-rolled-Metal strategies (ggml prefix: $GGML_PREFIX) =="
+
+  "$CC" -O2 -std=c11 -x objective-c \
+    -c eg_cosine_batch_strategy_metal_hand.m -o /tmp/eg_cosine_batch_strategy_metal_hand.o \
+    -framework Metal -framework Foundation
+
+  "$CC" -O2 -std=c11 -DEG_HAVE_STRATEGY_GGML -DEG_HAVE_STRATEGY_METAL_HAND -w \
+    -I"$GGML_PREFIX/include" \
+    vindex_bench.c engram_vindex.c \
+    eg_cosine_batch.c eg_cosine_batch_strategy_cpu.c eg_cosine_batch_strategy_ggml.c \
+    /tmp/eg_cosine_batch_strategy_metal_hand.o \
+    -L"$GGML_PREFIX/lib" -lggml -lggml-base \
+    -Wl,-rpath,"$GGML_PREFIX/lib" \
+    -lm -framework Metal -framework Foundation -o "$OUT"
+else
+  echo "== non-Darwin: building with the CPU-only fallback strategy (no ggml, no Metal) =="
+  "$CC" -O2 -std=c11 -w vindex_bench.c engram_vindex.c \
+    eg_cosine_batch.c eg_cosine_batch_strategy_cpu.c \
+    -lm -o "$OUT"
+fi
+
+echo "built: $OUT"

lang/runtime/eg_cosine_batch.c

@@ -0,0 +1,121 @@
+/* eg_cosine_batch.c — the Factory. Implements the stable public interface
+ * declared in eg_cosine_batch.h by selecting ONE concrete
+ * EgCosineBatchStrategy (eg_cosine_batch_strategy.h) and dispatching every
+ * call to it. This is the ONLY file that branches on EG_HAVE_STRATEGY_*
+ * (build-time: which strategy .c/.m files were actually compiled in for
+ * this platform) — call sites never see those macros.
+ *
+ * Selection is lazy (first call) and cached — mirrors the lazy-init caching
+ * every individual strategy already does internally, so there is no added
+ * per-call cost after the first.
+ *
+ * Selection mechanism (env var + build-time + runtime capability probe, all
+ * three, exactly as directed):
+ *   - BUILD-TIME decides which strategies exist to choose from at all: a
+ *     Darwin build compiles+links the ggml strategy and the hand-rolled
+ *     Metal strategy (EG_HAVE_STRATEGY_GGML / EG_HAVE_STRATEGY_METAL_HAND
+ *     both defined); a non-Darwin build compiles neither, matching PR #114's
+ *     original Linux behavior exactly (CPU-fallback only, no Objective-C
+ *     compiler or Metal frameworks required).
+ *   - RUNTIME CAPABILITY PROBE: each candidate strategy's own available()
+ *     does the real, cheap-after-first-call check (device present, backend
+ *     plugin loaded, pipeline compiled) — never assumed from build-time
+ *     alone. A build that HAS the ggml strategy compiled in but is running
+ *     on hardware/software where it can't actually initialize (backend
+ *     plugin missing, no GPU) correctly falls through to the next candidate.
+ *   - ENV VAR gives explicit, debuggable override for either axis:
+ *       EL_COSINE_BATCH_STRATEGY = "ggml" | "metal" | "cpu" | unset/"auto"
+ *         forces a specific strategy (falling back to cpu if the forced one
+ *         isn't actually available), or leaves the default auto-preference
+ *         order in place.
+ *       EL_METAL_COSINE = 0/n/N/f/F   (back-compat with PR #114's vindex_bench
+ *         gate) disables ALL GPU-backed strategies outright, same as before.
+ *
+ * DEFAULT preference order when nothing is forced: ggml, then hand-rolled
+ * Metal, then CPU fallback — first candidate whose available() reports true
+ * wins. This is what makes "stop hand-rolling GPU kernels, use ggml" real
+ * rather than nominal: ggml is what actually runs by default on this
+ * machine today (see the PR body for the measured numbers backing that).
+ */
+#include "eg_cosine_batch.h"
+#include "eg_cosine_batch_strategy.h"
+
+#include <stdlib.h>
+#include <string.h>
+
+static bool g_selected = false;
+static const EgCosineBatchStrategy* g_active = NULL;
+
+static bool eg_env_truthy_off(const char* v) {
+    return v && (v[0]=='0' || v[0]=='n' || v[0]=='N' || v[0]=='f' || v[0]=='F');
+}
+
+static const EgCosineBatchStrategy* eg_select_strategy(void) {
+    if (g_selected) return g_active;
+    g_selected = true;
+
+    const EgCosineBatchStrategy* cpu = eg_cosine_batch_strategy_cpu();
+    const char* force = getenv("EL_COSINE_BATCH_STRATEGY");
+    const char* legacy_off = getenv("EL_METAL_COSINE");
+
+    if (eg_env_truthy_off(legacy_off)) { g_active = cpu; return g_active; }
+
+    if (force && strcmp(force, "cpu") == 0) { g_active = cpu; return g_active; }
+
+    if (force && strcmp(force, "ggml") == 0) {
+#ifdef EG_HAVE_STRATEGY_GGML
+        const EgCosineBatchStrategy* s = eg_cosine_batch_strategy_ggml();
+        if (s->available()) { g_active = s; return g_active; }
+#endif
+        g_active = cpu; return g_active;
+    }
+
+    if (force && strcmp(force, "metal") == 0) {
+#ifdef EG_HAVE_STRATEGY_METAL_HAND
+        const EgCosineBatchStrategy* s = eg_cosine_batch_strategy_metal_hand();
+        if (s->available()) { g_active = s; return g_active; }
+#endif
+        g_active = cpu; return g_active;
+    }
+
+    /* auto (unset, or any other value): ggml -> metal-hand -> cpu, first
+     * available wins. */
+#ifdef EG_HAVE_STRATEGY_GGML
+    {
+        const EgCosineBatchStrategy* s = eg_cosine_batch_strategy_ggml();
+        if (s->available()) { g_active = s; return g_active; }
+    }
+#endif
+#ifdef EG_HAVE_STRATEGY_METAL_HAND
+    {
+        const EgCosineBatchStrategy* s = eg_cosine_batch_strategy_metal_hand();
+        if (s->available()) { g_active = s; return g_active; }
+    }
+#endif
+    g_active = cpu;
+    return g_active;
+}
+
+bool eg_cosine_batch_available(void) {
+    return eg_select_strategy()->available();
+}
+
+const char* eg_cosine_batch_strategy_name(void) {
+    return eg_select_strategy()->name;
+}
+
+bool eg_cosine_batch(const float* query, int32_t qdim,
+                      const float* const* node_ptrs,
+                      const int32_t* node_dims,
+                      int32_t n,
+                      double* out_scores) {
+    return eg_select_strategy()->batch(query, qdim, node_ptrs, node_dims, n, out_scores);
+}
+
+bool eg_cosine_batch_multi(const float* queries, int32_t qdim, int32_t nq,
+                            const float* const* node_ptrs,
+                            const int32_t* node_dims,
+                            int32_t n,
+                            double* out_scores) {
+    return eg_select_strategy()->batch_multi(queries, qdim, nq, node_ptrs, node_dims, n, out_scores);
+}

lang/runtime/eg_cosine_batch.h

@@ -0,0 +1,106 @@
+/* eg_cosine_batch.h — stable Adapter interface over batch-cosine-similarity
+ * BACKEND STRATEGIES. This header is the ONE thing call sites (el_runtime.c,
+ * vindex_bench.c, ...) talk to. Plain C11, safe to #include on every
+ * platform — the symbols declared here always exist and always link,
+ * regardless of what backend actually runs underneath. Zero #ifdef at call
+ * sites: which concrete strategy executes (ggml/Metal, hand-rolled Metal, or
+ * the always-false CPU fallback) is resolved once, lazily, inside
+ * eg_cosine_batch.c's factory — see eg_cosine_batch_strategy.h for that.
+ *
+ * This supersedes eg_metal_cosine.h (PR #114's single hand-rolled-Metal-only
+ * bridge). The contract is UNCHANGED — same shapes, same sentinel, same
+ * never-partial guarantee, same "caller must always be prepared to fall back
+ * to its own scalar per-node loop" rule — only the name changed, because the
+ * thing behind it is no longer "the Metal bridge," it is "whichever batch-
+ * cosine strategy the factory picked." eg_metal_cosine.h's original doc
+ * comments (byte-for-byte, this file is the direct descendant) are preserved
+ * below since they remain the precise spec any strategy must honor.
+ *
+ * On ANY failure at ANY step — no compute device, compile/init error, alloc
+ * failure, bad args — every function here returns false and writes nothing.
+ * Out-params are either fully populated or left completely untouched, never
+ * partial. Callers MUST always be prepared to fall back to their own scalar
+ * per-node CPU loop unconditionally. These functions must never crash, throw,
+ * or hang the calling process — several call sites run inside a long-lived
+ * daemon's request-handling hot path.
+ */
+#ifndef EG_COSINE_BATCH_H
+#define EG_COSINE_BATCH_H
+
+#include <stdint.h>
+#include <stdbool.h>
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/* Batched cosine similarity: one query vector against `n` node vectors.
+ *
+ *   query      — qdim floats, the query embedding. Raw/unnormalized.
+ *   qdim       — query dimensionality (e.g. 768 for nomic-embed-text).
+ *   node_ptrs  — array of n pointers, node_ptrs[i] pointing at a (possibly
+ *                differently-owned, possibly NULL) float vector for node i.
+ *                NOT required to be contiguous — every strategy performs the
+ *                gather into a packed row-major matrix internally, exactly
+ *                mirroring how EngramNode.emb is one malloc per node.
+ *   node_dims  — array of n ints, node_dims[i] = that node's real emb_dim
+ *                (0 or mismatched vs qdim => that node scores -2.0, matching
+ *                eg_cosine's null/dim-mismatch/zero-norm sentinel exactly).
+ *   n          — number of nodes.
+ *   out_scores — caller-owned array of n doubles; out_scores[i] is filled
+ *                with the cosine similarity of node i against query, or
+ *                -2.0 for a null/dim-mismatched/zero-norm node — bit-for-bit
+ *                the same contract as eg_cosine(node_ptrs[i], query, qdim).
+ *
+ * Returns true iff a real backend strategy ran and out_scores was fully
+ * populated. Returns false (out_scores left untouched) on ANY failure or
+ * unavailability — no compute device, compile/init failure, allocation
+ * failure, n<=0, qdim<=0, null query/node_ptrs/node_dims/out_scores.
+ */
+bool eg_cosine_batch(const float* query, int32_t qdim,
+                      const float* const* node_ptrs,
+                      const int32_t* node_dims,
+                      int32_t n,
+                      double* out_scores);
+
+/* True iff a real (non-CPU-fallback) strategy is available right now (cheap
+ * after the first call — cached). Purely informational (e.g. a startup log
+ * line or /api/stats field); callers should still treat a false return from
+ * eg_cosine_batch()/eg_cosine_batch_multi() itself as the authoritative
+ * fallback signal, not this function. */
+bool eg_cosine_batch_available(void);
+
+/* Which concrete strategy is currently selected — "ggml", "metal-hand",
+ * or "cpu-fallback". Purely informational/diagnostic, same spirit as
+ * eg_cosine_batch_available(). Never NULL. */
+const char* eg_cosine_batch_strategy_name(void);
+
+/* Multi-query batched cosine: nq query vectors against the SAME n node
+ * vectors, in one call. A real strategy uploads/prepares the node population
+ * ONCE and reuses it for every query, instead of nq separate
+ * eg_cosine_batch() calls each paying the full gather+upload cost — PR #114
+ * measured this necessary: at N~=13.7k/dim=768, repeating the single-query
+ * call per query was slower than the CPU baseline; batching queries together
+ * is what makes a GPU-backed path a real win at this shape. Use this
+ * whenever multiple queries will run against an unchanged (or
+ * rarely-changing) node population; use eg_cosine_batch() for a genuinely
+ * one-off comparison.
+ *
+ *   queries    — nq*qdim floats, row-major (query i at queries+i*qdim).
+ *   out_scores — caller-owned nq*n doubles, row-major
+ *                (out_scores[i*n+j] = cosine(queries[i], node j)), same
+ *                -2.0 sentinel semantics as eg_cosine_batch().
+ *
+ * Returns true iff a real strategy ran and out_scores was fully populated
+ * (all nq*n entries); false (untouched) on any failure/unavailability. */
+bool eg_cosine_batch_multi(const float* queries, int32_t qdim, int32_t nq,
+                            const float* const* node_ptrs,
+                            const int32_t* node_dims,
+                            int32_t n,
+                            double* out_scores);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif /* EG_COSINE_BATCH_H */

lang/runtime/eg_cosine_batch.metal

@@ -0,0 +1,156 @@
+/* eg_cosine_batch.metal — batched cosine similarity, one query vs N node vectors.
+ *
+ * GPU-shaped counterpart to eg_cosine() in el_runtime.c: same math, same
+ * dim-mismatch/zero-norm sentinel (-2.0), applied to N independent rows in
+ * parallel instead of one pair at a time in a CPU loop.
+ *
+ * Semantics MUST match eg_cosine() exactly:
+ *   - inputs are raw, UNNORMALIZED vectors (nomic-embed-text magnitudes are
+ *     not 1.0) — this kernel computes the full dot/(|a|*|b|) cosine, not a
+ *     plain dot product.
+ *   - a node whose declared dim differs from the query dim, or whose norm is
+ *     zero, scores exactly -2.0 (below any valid cosine in [-1,1]), so a
+ *     caller doing `if (score > threshold)` behaves identically whether the
+ *     scalar or the batched path filled the array.
+ *
+ * Precision: Apple GPUs do not support double in Metal Shading Language —
+ * everything here is float32. eg_cosine accumulates in CPU double, but its
+ * *inputs* are float32 embeddings, so the achievable precision ceiling is
+ * bounded by the input data regardless of accumulator width. To keep the
+ * float32 reduction from drifting relative to the double-accumulated CPU
+ * result across dim=768 terms, each thread accumulates with 4 independent
+ * partial sums (unrolled) rather than one running scalar — the same
+ * error-reduction trick already used by the CPU brute-force loop in
+ * vindex_bench.c. The measured float-vs-double delta is reported in the PR
+ * description; this is not assumed to be "close enough" without measurement.
+ */
+#include <metal_stdlib>
+using namespace metal;
+
+/* Per-dispatch invariants. `dim` is the query's dimensionality — the
+ * dimensionality every comparable node vector must match. */
+struct EgCosineParams {
+    uint n;      /* number of node rows */
+    uint dim;    /* vector width (both query and node rows are `dim` wide in
+                  * the packed buffer; node_dims[] carries each node's REAL
+                  * embedded dim for the mismatch check) */
+};
+
+/* One thread per node row. node_matrix is n*dim floats, row-major, packed at
+ * `dim` stride regardless of a row's real dim (the CPU side zero-pads or
+ * skips packing rows that don't match — see eg_cosine_batch_metal in
+ * eg_metal_cosine.m for the exact packing contract). node_dims[i] is the
+ * node's true emb_dim, used only for the mismatch sentinel — never used to
+ * index, since every row is packed at uniform `dim` stride. */
+kernel void eg_cosine_batch_kernel(
+    device const float*  query        [[buffer(0)]],
+    device const float*  node_matrix  [[buffer(1)]],
+    device const int*    node_dims    [[buffer(2)]],
+    constant EgCosineParams& p        [[buffer(3)]],
+    device float*        out_scores   [[buffer(4)]],
+    uint gid [[thread_position_in_grid]])
+{
+    if (gid >= p.n) return;
+
+    if (node_dims[gid] != int(p.dim)) {
+        out_scores[gid] = -2.0f;
+        return;
+    }
+
+    device const float* row = node_matrix + (uint64_t)gid * (uint64_t)p.dim;
+
+    /* 4-way partial accumulation — same shape as vindex_bench.c's brute_topk
+     * unroll, done here for float32 accuracy rather than raw throughput. */
+    float dot0 = 0.0f, dot1 = 0.0f, dot2 = 0.0f, dot3 = 0.0f;
+    float na0  = 0.0f, na1  = 0.0f, na2  = 0.0f, na3  = 0.0f;
+    float nb0  = 0.0f, nb1  = 0.0f, nb2  = 0.0f, nb3  = 0.0f;
+
+    uint d = 0;
+    uint dim4 = p.dim & ~3u;
+    for (; d < dim4; d += 4) {
+        float a0 = row[d],   b0 = query[d];
+        float a1 = row[d+1], b1 = query[d+1];
+        float a2 = row[d+2], b2 = query[d+2];
+        float a3 = row[d+3], b3 = query[d+3];
+        dot0 += a0*b0; dot1 += a1*b1; dot2 += a2*b2; dot3 += a3*b3;
+        na0  += a0*a0; na1  += a1*a1; na2  += a2*a2; na3  += a3*a3;
+        nb0  += b0*b0; nb1  += b1*b1; nb2  += b2*b2; nb3  += b3*b3;
+    }
+    float dot = (dot0 + dot1) + (dot2 + dot3);
+    float na  = (na0  + na1)  + (na2  + na3);
+    float nb  = (nb0  + nb1)  + (nb2  + nb3);
+    for (; d < p.dim; d++) {
+        float a = row[d], b = query[d];
+        dot += a*b; na += a*a; nb += b*b;
+    }
+
+    if (na <= 0.0f || nb <= 0.0f) {
+        out_scores[gid] = -2.0f;
+        return;
+    }
+    out_scores[gid] = dot / sqrt(na * nb);
+}
+
+/* ── multi-query variant ──────────────────────────────────────────────────
+ * Same per-pair math as eg_cosine_batch_kernel, but amortizes ONE upload of
+ * node_matrix (the expensive part at real store size — 13k*768 floats is
+ * ~42MB) across `nq` queries instead of re-uploading it once per query.
+ * Measured need: a naive one-query-at-a-time loop calling the single-query
+ * kernel nq times was SLOWER than the CPU oracle at N≈13.7k (re-gather +
+ * re-upload dominated the actual compute) — this is the fix, not a
+ * hypothetical optimization.
+ *
+ * 2D grid: x = node index [0,n), y = query index [0,nq). out_scores is
+ * nq*n, row-major by query (out_scores[qid*n + nid]). */
+struct EgCosineMultiParams { uint n; uint dim; uint nq; };
+
+kernel void eg_cosine_batch_multi_kernel(
+    device const float*  queries      [[buffer(0)]],   /* nq*dim */
+    device const float*  node_matrix  [[buffer(1)]],   /* n*dim */
+    device const int*    node_dims    [[buffer(2)]],   /* n */
+    constant EgCosineMultiParams& p   [[buffer(3)]],
+    device float*        out_scores   [[buffer(4)]],   /* nq*n */
+    uint2 gid [[thread_position_in_grid]])
+{
+    uint nid = gid.x, qid = gid.y;
+    if (nid >= p.n || qid >= p.nq) return;
+
+    uint64_t out_idx = (uint64_t)qid * (uint64_t)p.n + (uint64_t)nid;
+
+    if (node_dims[nid] != int(p.dim)) {
+        out_scores[out_idx] = -2.0f;
+        return;
+    }
+
+    device const float* row   = node_matrix + (uint64_t)nid * (uint64_t)p.dim;
+    device const float* query = queries     + (uint64_t)qid * (uint64_t)p.dim;
+
+    float dot0 = 0.0f, dot1 = 0.0f, dot2 = 0.0f, dot3 = 0.0f;
+    float na0  = 0.0f, na1  = 0.0f, na2  = 0.0f, na3  = 0.0f;
+    float nb0  = 0.0f, nb1  = 0.0f, nb2  = 0.0f, nb3  = 0.0f;
+
+    uint d = 0;
+    uint dim4 = p.dim & ~3u;
+    for (; d < dim4; d += 4) {
+        float a0 = row[d],   b0 = query[d];
+        float a1 = row[d+1], b1 = query[d+1];
+        float a2 = row[d+2], b2 = query[d+2];
+        float a3 = row[d+3], b3 = query[d+3];
+        dot0 += a0*b0; dot1 += a1*b1; dot2 += a2*b2; dot3 += a3*b3;
+        na0  += a0*a0; na1  += a1*a1; na2  += a2*a2; na3  += a3*a3;
+        nb0  += b0*b0; nb1  += b1*b1; nb2  += b2*b2; nb3  += b3*b3;
+    }
+    float dot = (dot0 + dot1) + (dot2 + dot3);
+    float na  = (na0  + na1)  + (na2  + na3);
+    float nb  = (nb0  + nb1)  + (nb2  + nb3);
+    for (; d < p.dim; d++) {
+        float a = row[d], b = query[d];
+        dot += a*b; na += a*a; nb += b*b;
+    }
+
+    if (na <= 0.0f || nb <= 0.0f) {
+        out_scores[out_idx] = -2.0f;
+        return;
+    }
+    out_scores[out_idx] = dot / sqrt(na * nb);
+}

lang/runtime/eg_cosine_batch_strategy.h

@@ -0,0 +1,82 @@
+/* eg_cosine_batch_strategy.h — internal Strategy interface, NOT for call
+ * sites (they use eg_cosine_batch.h). Only eg_cosine_batch.c's factory and
+ * the concrete strategy implementation files include this.
+ *
+ * Each concrete strategy exposes exactly one getter returning a pointer to a
+ * static, immutable EgCosineBatchStrategy vtable. Which getters actually
+ * exist as linkable symbols is a BUILD-TIME concern (decided by
+ * build_vindex_bench.sh / the engram daemon's own build, via which .c/.m
+ * files get compiled per platform) gated by the EG_HAVE_STRATEGY_* macros
+ * below — the factory in eg_cosine_batch.c is the ONLY place that branches
+ * on those macros. Call sites never see them; that's the whole point of the
+ * Adapter in eg_cosine_batch.h.
+ *
+ * Three concrete strategies exist:
+ *   eg_cosine_batch_strategy_ggml()        — ggml + dynamically-loaded Metal
+ *                                             backend plugin. Darwin only in
+ *                                             this build; the default
+ *                                             preferred strategy wherever
+ *                                             available. EG_HAVE_STRATEGY_GGML.
+ *   eg_cosine_batch_strategy_metal_hand()  — the original hand-rolled Metal
+ *                                             compute shader from PR #114
+ *                                             (eg_cosine_batch.metal),
+ *                                             preserved verbatim as a
+ *                                             selectable fallback strategy,
+ *                                             not deleted. Darwin only.
+ *                                             EG_HAVE_STRATEGY_METAL_HAND.
+ *   eg_cosine_batch_strategy_cpu()         — universal always-false
+ *                                             fallback. Always compiled, on
+ *                                             every platform; this is what a
+ *                                             non-Darwin build links
+ *                                             exclusively (matching PR #114's
+ *                                             eg_metal_cosine_stub.c), and
+ *                                             what any platform falls back
+ *                                             to when no real strategy is
+ *                                             available at runtime.
+ */
+#ifndef EG_COSINE_BATCH_STRATEGY_H
+#define EG_COSINE_BATCH_STRATEGY_H
+
+#include <stdint.h>
+#include <stdbool.h>
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef struct EgCosineBatchStrategy {
+    /* Stable, short, lowercase-hyphenated identifier — what
+     * eg_cosine_batch_strategy_name() surfaces. Never NULL. */
+    const char* name;
+
+    /* Cheap after the first call (lazy init, cached internally). Must never
+     * throw/crash/hang — mirrors eg_cosine_batch_available()'s contract. */
+    bool (*available)(void);
+
+    /* Same shape/contract as eg_cosine_batch() in eg_cosine_batch.h. */
+    bool (*batch)(const float* query, int32_t qdim,
+                  const float* const* node_ptrs, const int32_t* node_dims,
+                  int32_t n, double* out_scores);
+
+    /* Same shape/contract as eg_cosine_batch_multi() in eg_cosine_batch.h. */
+    bool (*batch_multi)(const float* queries, int32_t qdim, int32_t nq,
+                         const float* const* node_ptrs, const int32_t* node_dims,
+                         int32_t n, double* out_scores);
+} EgCosineBatchStrategy;
+
+#ifdef EG_HAVE_STRATEGY_GGML
+const EgCosineBatchStrategy* eg_cosine_batch_strategy_ggml(void);
+#endif
+
+#ifdef EG_HAVE_STRATEGY_METAL_HAND
+const EgCosineBatchStrategy* eg_cosine_batch_strategy_metal_hand(void);
+#endif
+
+/* Always declared/linked, on every platform/build. */
+const EgCosineBatchStrategy* eg_cosine_batch_strategy_cpu(void);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif /* EG_COSINE_BATCH_STRATEGY_H */

lang/runtime/eg_cosine_batch_strategy_cpu.c

@@ -0,0 +1,45 @@
+/* eg_cosine_batch_strategy_cpu.c — plain-C, zero-dependency universal
+ * fallback strategy. Always returns false / unavailable. Direct descendant
+ * of PR #114's eg_metal_cosine_stub.c, generalized from "the Metal stub" to
+ * "the strategy vtable's universal fallback entry" now that multiple real
+ * strategies can exist.
+ *
+ * Always compiled, on every platform. On Darwin builds it is the last-resort
+ * strategy the factory falls back to when neither ggml nor the hand-rolled
+ * Metal strategy is available at runtime (no device, compile failure, ...).
+ * On non-Darwin builds it is the ONLY strategy compiled in at all — no
+ * Objective-C, no Metal frameworks, no ggml/Metal backend plugin — so
+ * eg_cosine_batch()/eg_cosine_batch_multi() always return false there and
+ * every call site's existing CPU fallback runs unconditionally, exactly as
+ * before this PR.
+ */
+#include "eg_cosine_batch_strategy.h"
+
+static bool cpu_available(void) {
+    return false;
+}
+
+static bool cpu_batch(const float* query, int32_t qdim,
+                       const float* const* node_ptrs, const int32_t* node_dims,
+                       int32_t n, double* out_scores) {
+    (void)query; (void)qdim; (void)node_ptrs; (void)node_dims; (void)n; (void)out_scores;
+    return false;
+}
+
+static bool cpu_batch_multi(const float* queries, int32_t qdim, int32_t nq,
+                             const float* const* node_ptrs, const int32_t* node_dims,
+                             int32_t n, double* out_scores) {
+    (void)queries; (void)qdim; (void)nq; (void)node_ptrs; (void)node_dims; (void)n; (void)out_scores;
+    return false;
+}
+
+static const EgCosineBatchStrategy g_cpu_strategy = {
+    .name = "cpu-fallback",
+    .available = cpu_available,
+    .batch = cpu_batch,
+    .batch_multi = cpu_batch_multi,
+};
+
+const EgCosineBatchStrategy* eg_cosine_batch_strategy_cpu(void) {
+    return &g_cpu_strategy;
+}

lang/runtime/eg_cosine_batch_strategy_ggml.c

@@ -0,0 +1,515 @@
+/* eg_cosine_batch_strategy_ggml.c — the GGML Strategy, and the preferred
+ * default whenever it is available (see the factory's selection order in
+ * eg_cosine_batch.c).
+ *
+ * WHY: directive from Will Anderson — stop hand-rolling GPU kernels, use a
+ * real, proven, permissively-licensed library instead. ggml (the compute
+ * library underneath llama.cpp, MIT licensed) is already installed on this
+ * machine as a standalone Homebrew package (`brew info ggml`), independent
+ * of llama.cpp itself. This file is a genuinely bounded COMPUTE UTILITY —
+ * batch cosine-similarity math — analogous to a VBD Accessor calling out to
+ * infrastructure. It is explicitly NOT the engram's reasoning/persistence
+ * core; using ggml here does not cross the "own the core" line, because
+ * batch cosine math is infrastructure, not the graph traversal / activation
+ * spreading / "thinking" that IS the core and stays 100% own-code.
+ *
+ * ── The real API shape (verified against the installed headers + a
+ *    standalone probe program, not assumed from memory of other tensor
+ *    libraries) ──────────────────────────────────────────────────────────
+ *
+ * ggml ships its CPU and Metal implementations as DYNAMICALLY LOADED PLUGIN
+ * .so files (confirmed by nm: `ggml_backend_metal_init` is NOT an exported
+ * symbol of libggml.dylib/libggml-base.dylib — it exists ONLY inside
+ * libggml-metal.so under $(brew --prefix ggml)/libexec/). You cannot link
+ * `-lggml-metal`; you must go through ggml's backend REGISTRY:
+ *
+ *   1. ggml_backend_load_all_from_path(dir) — dlopen()s every backend plugin
+ *      .so found in `dir` and registers its device(s). We point this at
+ *      $(brew --prefix ggml)/libexec (resolved once, at build+init time; see
+ *      eg_ggml_backend_dir() below) rather than relying on
+ *      ggml_backend_load_all()'s own default search heuristics, which are
+ *      tuned for an installed llama.cpp-style app bundle layout, not an
+ *      arbitrary `cc`-built binary invoked from an arbitrary cwd — the exact
+ *      same "must not silently fall back to CPU for reasons that have
+ *      nothing to do with GPU availability" concern PR #114's hand-rolled
+ *      bridge already documented for its own embedded-shader-source choice.
+ *   2. ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_GPU) — find the
+ *      registered Metal device.
+ *   3. ggml_backend_dev_init(dev, NULL) — get a live ggml_backend_t.
+ *   4. Build a tiny ggml_context (no_alloc=true; it holds only tensor
+ *      metadata, not data), declare 2D F32 tensors, ggml_mul_mat(nodes,
+ *      query) — ggml's documented convention: A is [k cols, n rows], B is
+ *      [k cols, m rows] (transposed internally), result is [n cols, m rows]
+ *      — i.e. mul_mat(node_matrix[dim,n], query_matrix[dim,nq]) yields
+ *      out[n,nq] where out[j*n+i] = dot(node_i, query_j). A row-major
+ *      (dim,n) node matrix and a row-major (dim,nq) query matrix is EXACTLY
+ *      the packed layout the hand-rolled Metal kernel already used — one
+ *      matmul replaces the whole per-row dot-product loop.
+ *   5. ggml_backend_alloc_ctx_tensors(ctx, backend) to actually allocate
+ *      device buffers for those tensors, ggml_backend_tensor_set() to upload,
+ *      ggml_backend_graph_compute() to run, ggml_backend_tensor_get() to
+ *      read back.
+ *
+ * This exact sequence was verified end-to-end in a standalone probe (build
+ * it yourself: see the PR description) against a plain-C CPU dot product —
+ * bit-for-bit correct within float rounding. Real numbers against the
+ * el_runtime.c CPU oracle are reported in the PR body via vindex_bench.
+ *
+ * ggml_mul_mat only computes the raw dot products — it has no notion of
+ * "cosine" or of this codebase's -2.0 dim-mismatch/null/zero-norm sentinel.
+ * Per the adapter's directive: gather only VALID, uniform-dim rows into the
+ * packed matrix sent to the GPU (skipping null/mismatched rows entirely,
+ * rather than the hand-rolled kernel's zero-pad-and-sentinel-in-shader
+ * approach), then scatter -2.0 back for every row that was excluded — same
+ * gather/scatter contract eg_cosine_batch.h documents. Norms (||node||,
+ * ||query||) are computed on the CPU host in the same pass that already
+ * touches every element to gather/convert — essentially free — using the
+ * same 4-way-partial-sum accumulation the hand-rolled kernel and the CPU
+ * oracle both use, so the float32 error profile stays comparable across all
+ * three strategies. Only the O(n*dim*nq) dot-product matmul — the actual
+ * expensive part — is offloaded to the GPU.
+ *
+ * ── Precision: the ne11<=8 chunking, and why it is not optional ──────────
+ *
+ * The claim in the first version of this file — "ggml_mul_mat on F32 x F32
+ * inputs computes in F32 on the Metal backend" — is WRONG, and the 0.9933
+ * id-recall it shipped with (vs the hand-rolled kernel's 0.9997) was the
+ * symptom. ggml-metal has two F32xF32 matmul kernels and picks between them
+ * purely on ne11 (the number of B rows == our query count):
+ *
+ *   ne11 <= 8  -> kernel_mul_mv_ext_f32_f32_* / kernel_mul_mv_f32_f32_*
+ *                 templated <float, float> — genuine F32 accumulation.
+ *   ne11 >  8  -> kernel_mul_mm_f32_f32, which is templated
+ *                 <half, half4x4, simdgroup_half8x8, half, half2x4,
+ *                  simdgroup_half8x8, ...> — i.e. BOTH operands are narrowed
+ *                 to F16 and accumulated in simdgroup_half8x8 tiles, even
+ *                 though the tensors are GGML_TYPE_F32 on both sides.
+ *
+ * (Read it yourself, no guessing — the kernel templates are literal strings
+ * in the shipped plugin:
+ *    strings $(brew --prefix ggml)/libexec/libggml-metal.so \
+ *      | grep -E 'host_name\("kernel_mul_m[mv]_f32_f32'
+ *  and the runtime pick is visible with GGML_METAL_DEBUG-style logging as
+ *  "compiling pipeline: base = 'kernel_mul_mm_f32_f32'".)
+ *
+ * The previous code issued ONE ggml_mul_mat with ne11 = nq (300 in the
+ * benchmark), landing squarely on the F16 mul_mm path. Measured on this
+ * machine (M4 Pro), n=13415 x dim=768 x nq=300, against a CPU double-
+ * accumulated oracle:
+ *
+ *   ne11=300 (one mul_mat, the old code) : mean |Δdot| = 1.038e-05
+ *   ne11=8   (chunked, this code)        : mean |Δdot| = 3.863e-09
+ *
+ * — a ~2700x reduction in dot-product error, which is exactly the gap that
+ * showed up as 0.9933-vs-0.9997 recall.
+ *
+ * ggml_mul_mat_set_prec(t, GGML_PREC_F32) does NOT fix this. It was tried:
+ * the error was bit-identical with and without it (1.038e-05 either way),
+ * because ggml-metal only consults the prec flag on paths that have an F32
+ * variant to switch to, and there is no F32-accumulating mul_mm kernel in
+ * this build to select. The ONLY lever from outside ggml is ne11.
+ *
+ * So: instead of one mul_mat with ne11=nq, we emit ceil(nq/8) mul_mats, each
+ * over an ne11<=8 ggml_view_2d slice of the same query tensor, all into ONE
+ * graph and ONE ggml_backend_graph_compute. The node matrix is still uploaded
+ * exactly once and still read by the GPU as one shared operand — the whole
+ * point of batch_multi is preserved.
+ *
+ * The cost is real, and stated rather than buried. Timing the whole
+ * batch_multi() call (gather + norms + upload + GPU + scatter) on the real
+ * shape, median of 15 reps after a discarded warm-up, three separate runs:
+ *
+ *   unchunked (old, F16 mm) : 13.19 / 13.35 / 14.42 ms   -> ~0.044 ms/query
+ *   chunked   (this code)   : 19.92 / 20.08 / 20.23 ms   -> ~0.067 ms/query
+ *   hand-rolled Metal       : 17.74 / 17.88 / 17.99 ms   -> ~0.060 ms/query
+ *
+ * So correctness here costs about +6.7ms per 300-query batch (~1.5x on this
+ * call), and leaves us ~12% behind the hand-rolled kernel instead of ~35%
+ * ahead of it. That is not free and should not be sold as free. The reason it
+ * cannot be recovered inside ggml: an fp32 matmul on Metal has to re-stream
+ * the whole node matrix once per <=8 queries (38 dispatches x ~41MB here),
+ * where the F16 mul_mm kernel tiles it in threadgroup memory and reads it far
+ * fewer times. ggml's Metal backend ships no fp32 TILED matmul, so on this
+ * backend "fast" and "fp32" are genuinely exclusive — the hand-rolled kernel
+ * escapes the choice only because it is an fp32 kernel written for this one
+ * shape. Trading precision back for speed is a one-line env change; trading
+ * the other way was not available before this commit at all.
+ *
+ * 8 is not a magic number we invented — it is ggml-metal's own mul_mm
+ * threshold, measured by sweeping ne11 and watching both the error and which
+ * pipeline ggml compiles (9 flips to mul_mm and the error jumps back to
+ * 1.0e-05 in the same step). EL_GGML_MULMAT_CHUNK overrides it: raise it to
+ * trade this precision back for throughput, or set it >= nq to reproduce the
+ * old single-mul_mat behaviour exactly. If a future ggml moves the threshold,
+ * the worst case is that we silently land back on mul_mm — the same accuracy
+ * we shipped before, never a correctness break.
+ *
+ * ── Cold start: what is and is not ours to fix ───────────────────────────
+ *
+ * The ~7.8s first-call cost reported for the first version of this file is
+ * NOT this file re-initialising per call (init is, and always was, cached
+ * behind g_init_attempted below). It is Apple's Metal shader cache missing
+ * on ggml's embedded metallib — ggml-metal ships ~650 kernels in one
+ * __ggml_metallib section, and the first newLibraryWithData of it on a given
+ * machine costs seconds ("ggml_metal_library_init: loaded in 7.670 sec")
+ * while the driver populates ~/…/C/com.apple.metal/. That cache is keyed on
+ * the library, not on our binary, and is shared across processes: the very
+ * next run of a DIFFERENT binary linking the same ggml reports
+ * "loaded in 0.009 sec". So it is a once-per-machine, per-ggml-version cost,
+ * not a per-process one, and nothing this file does can avoid it — the
+ * hand-rolled strategy escapes it only because its shader is two small
+ * kernels instead of six hundred.
+ *
+ * The residual warm init IS ours to look at, and the answer there is "there
+ * was nothing much to win": ggml_backend_load_all_from_path() dlopens every
+ * plugin in the directory (three CPU micro-arch variants + BLAS + Metal) when
+ * we only ever use Metal, so we now load the single Metal plugin instead —
+ * but measured warm that is 44.7-52.4ms against 46.9-58.9ms, i.e. the same
+ * number inside noise, because libggml-metal.so's own init dominates. Warm
+ * ggml init lands at 44-53ms, against 36-117ms for the hand-rolled strategy's
+ * device+pipeline setup. Cold start was never the real defect here; precision
+ * was.
+ */
+#include "eg_cosine_batch_strategy.h"
+
+#include <ggml.h>
+#include <ggml-backend.h>
+#include <ggml-alloc.h>
+
+#include <stdint.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <math.h>
+
+/* ── lazy, one-time backend init, cached ─────────────────────────────────── */
+static bool g_init_attempted = false;
+static bool g_init_ok = false;
+static ggml_backend_t g_backend = NULL;
+
+/* Where to look for the dynamically-loaded backend plugin .so files.
+ * EL_GGML_BACKEND_PATH overrides for non-standard installs; otherwise we try
+ * the Homebrew opt-prefix symlink (stable across ggml point-version bumps —
+ * $(brew --prefix ggml)/libexec — confirmed to exist and contain
+ * libggml-metal.so / libggml-cpu-*.so / libggml-blas.so on this machine),
+ * falling back to ggml's own default search (ggml_backend_load_all()) in
+ * case a different install layout (e.g. a from-source build with a
+ * standard-prefix install) makes that succeed instead. */
+static const char* eg_ggml_backend_dir(void) {
+    const char* s = getenv("EL_GGML_BACKEND_PATH");
+    if (s && *s) return s;
+    return "/opt/homebrew/opt/ggml/libexec";
+}
+
+/* "<dir>/libggml-metal.so" in a static buffer. Only ever called once, from
+ * eg_ggml_ensure_init(), before any thread could race it. */
+static const char* eg_ggml_metal_plugin_path(const char* dir) {
+    static char buf[1024];
+    snprintf(buf, sizeof buf, "%s/libggml-metal.so", dir);
+    return buf;
+}
+
+/* Largest ne11 (query-batch rows per ggml_mul_mat) that keeps ggml-metal on
+ * its F32 mul_mv kernels instead of the F16-accumulating mul_mm kernel — see
+ * the precision discussion in this file's header. EL_GGML_MULMAT_CHUNK
+ * overrides; a value <= 0 means "use the default". */
+#define EG_GGML_MULMAT_CHUNK_DEFAULT 8
+
+static int32_t eg_ggml_mulmat_chunk(void) {
+    static bool resolved = false;
+    static int32_t chunk = EG_GGML_MULMAT_CHUNK_DEFAULT;
+    if (!resolved) {
+        resolved = true;
+        const char* s = getenv("EL_GGML_MULMAT_CHUNK");
+        if (s && *s) {
+            long v = strtol(s, NULL, 10);
+            if (v > 0 && v <= INT32_MAX) chunk = (int32_t)v;
+        }
+    }
+    return chunk;
+}
+
+/* Which ggml device this strategy computes on. GPU (Metal) is the default
+ * because offloading is the architectural point — the engram's own graph
+ * traversal and activation spreading are CPU work, and a "GPU" strategy that
+ * quietly saturates the CPU steals from them.
+ *
+ * ACCEL (ggml's BLAS/Accelerate plugin) is reachable here mainly as a
+ * portability fallback and a diagnostic, and it is documented as MEASURED AND
+ * REJECTED rather than as a recommendation. In an isolated probe that timed
+ * only ggml_backend_graph_compute, BLAS looked excellent — 3.4-4.0ms for the
+ * 300-query batch at mean |Δdot| 1.5e-08, i.e. as fast as the old F16 path and
+ * far more accurate. End to end on the real store through vindex_bench it does
+ * not hold up: 0.191 ms/query at id-recall 0.9973, against 0.125-0.142 ms/query
+ * at 0.9987 for the Metal default. It is dominated on BOTH axes, because the
+ * isolated probe was not competing with the rest of the batch for the same CPU
+ * cores and the real call path is. Kept because a machine with no usable Metal
+ * device still wants a working ggml strategy — not because it is faster. */
+static enum ggml_backend_dev_type eg_ggml_device_type(void) {
+    const char* s = getenv("EL_GGML_DEVICE");
+    if (s && *s) {
+        if (strcmp(s, "accel") == 0) return GGML_BACKEND_DEVICE_TYPE_ACCEL;
+        if (strcmp(s, "cpu")   == 0) return GGML_BACKEND_DEVICE_TYPE_CPU;
+    }
+    return GGML_BACKEND_DEVICE_TYPE_GPU;
+}
+
+static bool eg_ggml_ensure_init(void) {
+    if (g_init_attempted) return g_init_ok;
+    g_init_attempted = true;
+
+    const char* dir = eg_ggml_backend_dir();
+    const enum ggml_backend_dev_type want = eg_ggml_device_type();
+
+    /* Metal is the only backend this strategy uses by default, so load just
+     * that one plugin rather than dlopening the whole directory (three CPU
+     * micro-arch variants + BLAS + Metal here).
+     *
+     * Be honest about what this buys: almost nothing in wall time. Measured
+     * warm, three runs each — load-everything 58.9/46.9/55.1ms, Metal-only
+     * 52.4/51.2/44.7ms. The cost is dominated by dlopening and initialising
+     * libggml-metal.so itself, not by the four plugins we skip, so the two
+     * overlap inside noise. It is kept because registering four device types
+     * we will never dispatch to is untidy and makes ggml_backend_dev_by_type
+     * ambiguous, not because it is a speedup — do not cite it as one.
+     *
+     * ggml_backend_load() returns NULL for a missing or unloadable path,
+     * which simply falls through to the broader searches below; it is never
+     * fatal. Any non-default device needs the full directory scan to find
+     * its plugin, so skip the fast path there. */
+    if (want == GGML_BACKEND_DEVICE_TYPE_GPU)
+        ggml_backend_load(eg_ggml_metal_plugin_path(dir));
+
+    ggml_backend_dev_t dev = ggml_backend_dev_by_type(want);
+    if (!dev) {
+        /* Non-standard layout, a ggml built with a differently-named Metal
+         * plugin, or a non-default device: dlopen every plugin in `dir`. */
+        ggml_backend_load_all_from_path(dir);
+        dev = ggml_backend_dev_by_type(want);
+    }
+    if (!dev) {
+        /* Fall back to ggml's own default search heuristics only if the
+         * explicit path above found nothing — avoids double-registering the
+         * same plugins (ggml does not dedupe two different paths that
+         * happen to resolve to the same files, e.g. our stable opt-prefix
+         * symlink vs. its own Cellar-relative guess) in the common case
+         * where the explicit path already worked. */
+        ggml_backend_load_all();
+        dev = ggml_backend_dev_by_type(want);
+    }
+    if (!dev) return false;
+
+    ggml_backend_t backend = ggml_backend_dev_init(dev, NULL);
+    if (!backend) return false;
+
+    g_backend = backend;
+    g_init_ok = true;
+    return true;
+}
+
+static bool ggml_strategy_available(void) {
+    return eg_ggml_ensure_init();
+}
+
+/* ── shared core: gather valid rows + norms, matmul, scatter ────────────── */
+
+/* 4-way partial-sum squared-norm accumulation over `dim` floats — same shape
+ * as eg_cosine_batch.metal's per-thread accumulation and vindex_bench.c's
+ * CPU brute_topk unroll, kept consistent on purpose so the float32 error
+ * profile is comparable across all three strategies. */
+static float eg_norm_sq_f32(const float* v, int32_t dim) {
+    float s0 = 0, s1 = 0, s2 = 0, s3 = 0;
+    int32_t d = 0, dim4 = dim & ~3;
+    for (; d < dim4; d += 4) {
+        s0 += v[d] * v[d]; s1 += v[d+1] * v[d+1];
+        s2 += v[d+2] * v[d+2]; s3 += v[d+3] * v[d+3];
+    }
+    float s = (s0 + s1) + (s2 + s3);
+    for (; d < dim; d++) s += v[d] * v[d];
+    return s;
+}
+
+/* Runs one ggml_mul_mat(node_matrix[dim,n_valid], query_matrix[dim,nq]) and
+ * combines it with CPU-computed norms into cosine scores, scattering into
+ * out_scores at ORIGINAL (ungathered) indices. out_scores must already be
+ * fully sized for n*nq (or n for the single-query case, nq=1) — every entry
+ * gets written (valid rows get a real cosine, invalid rows get -2.0), so
+ * this never leaves a partial result. Returns false only on a genuine
+ * failure (alloc, compute) at which point out_scores is left as whatever a
+ * caller-supplied scratch buffer already contained — callers here always
+ * pass a fresh buffer they discard on false, matching the adapter contract
+ * of "on failure, out_scores is treated as untouched" from the caller's
+ * point of view. */
+static bool eg_ggml_run(const float* queries, int32_t qdim, int32_t nq,
+                         const float* const* node_ptrs, const int32_t* node_dims,
+                         int32_t n, double* out_scores) {
+    if (!queries || qdim <= 0 || nq <= 0 || !node_ptrs || !node_dims || n <= 0 || !out_scores)
+        return false;
+    if (!eg_ggml_ensure_init()) return false;
+
+    /* Pass 1 (CPU): gather valid rows (non-NULL ptr, dim == qdim) into a
+     * packed (dim, n_valid) row-major matrix, remembering the original index
+     * of each packed row, and compute each valid row's squared norm in the
+     * same pass. Rows excluded here get -2.0 scattered for every query
+     * below without ever touching the GPU. */
+    int32_t* valid_orig = (int32_t*)malloc((size_t)n * sizeof(int32_t));
+    float* node_norm_sq  = (float*)malloc((size_t)n * sizeof(float)); /* indexed by packed position */
+    float* node_matrix   = NULL;
+    if (!valid_orig || !node_norm_sq) { free(valid_orig); free(node_norm_sq); return false; }
+
+    int32_t n_valid = 0;
+    for (int32_t i = 0; i < n; i++) {
+        if (node_ptrs[i] && node_dims[i] == qdim) n_valid++;
+    }
+
+    if (n_valid > 0) {
+        node_matrix = (float*)malloc((size_t)n_valid * (size_t)qdim * sizeof(float));
+        if (!node_matrix) { free(valid_orig); free(node_norm_sq); return false; }
+        int32_t w = 0;
+        for (int32_t i = 0; i < n; i++) {
+            if (!node_ptrs[i] || node_dims[i] != qdim) continue;
+            memcpy(node_matrix + (size_t)w * qdim, node_ptrs[i], (size_t)qdim * sizeof(float));
+            node_norm_sq[w] = eg_norm_sq_f32(node_ptrs[i], qdim);
+            valid_orig[w] = i;
+            w++;
+        }
+    }
+
+    /* Query norms — nq is typically small (1 or the size of one batch of
+     * comparison queries), so this loop is cheap regardless. */
+    float* q_norm_sq = (float*)malloc((size_t)nq * sizeof(float));
+    if (!q_norm_sq) { free(valid_orig); free(node_norm_sq); free(node_matrix); return false; }
+    for (int32_t j = 0; j < nq; j++) q_norm_sq[j] = eg_norm_sq_f32(queries + (size_t)j * qdim, qdim);
+
+    /* Nothing valid to compare against: every output is -2.0. Still a fully
+     * and correctly populated result — no GPU dispatch was needed to know
+     * that. */
+    if (n_valid == 0) {
+        for (size_t k = 0; k < (size_t)n * (size_t)nq; k++) out_scores[k] = -2.0;
+        free(valid_orig); free(node_norm_sq); free(node_matrix); free(q_norm_sq);
+        return true;
+    }
+
+    /* Pass 2 (GPU via ggml): dot[j*n_valid + i] = dot(node_i, query_j),
+     * computed as ceil(nq/chunk) separate ggml_mul_mat ops over ne11<=chunk
+     * ggml_view_2d slices of ONE query tensor, all expanded into ONE graph
+     * and run by ONE ggml_backend_graph_compute. Chunking is what keeps
+     * ggml-metal on its F32 mul_mv kernels rather than the F16-accumulating
+     * mul_mm kernel (see this file's header); sharing one graph and one
+     * t_nodes tensor is what keeps the node matrix uploaded exactly once,
+     * which is the entire reason batch_multi exists. */
+    const int32_t chunk = eg_ggml_mulmat_chunk();
+    const int32_t ngroups = (nq + chunk - 1) / chunk;
+
+    /* Tensors held by the context: t_nodes, t_query, plus one view and one
+     * mul_mat result per group. The graph holds at most one node per view and
+     * one per mul_mat. Slack on both so a ggml that bookkeeps slightly
+     * differently cannot silently overflow the arena. */
+    const size_t n_tensors  = (size_t)2 * (size_t)ngroups + 8;
+    const size_t graph_size = (size_t)2 * (size_t)ngroups + 16;
+    struct ggml_init_params gp = {
+        .mem_size   = ggml_tensor_overhead() * n_tensors
+                    + ggml_graph_overhead_custom(graph_size, false),
+        .mem_buffer = NULL,
+        .no_alloc   = true,
+    };
+    struct ggml_context* ctx = ggml_init(gp);
+    if (!ctx) { free(valid_orig); free(node_norm_sq); free(node_matrix); free(q_norm_sq); return false; }
+
+    struct ggml_tensor** t_dots = (struct ggml_tensor**)malloc((size_t)ngroups * sizeof(*t_dots));
+    if (!t_dots) { ggml_free(ctx); free(valid_orig); free(node_norm_sq); free(node_matrix); free(q_norm_sq); return false; }
+
+    struct ggml_tensor* t_nodes = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, qdim, n_valid);
+    struct ggml_tensor* t_query = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, qdim, nq);
+    struct ggml_cgraph*  gf     = t_nodes && t_query
+                                ? ggml_new_graph_custom(ctx, graph_size, false) : NULL;
+    if (!gf) { free(t_dots); ggml_free(ctx); free(valid_orig); free(node_norm_sq); free(node_matrix); free(q_norm_sq); return false; }
+
+    bool built = true;
+    for (int32_t g = 0; g < ngroups; g++) {
+        const int32_t start = g * chunk;
+        const int32_t count = (start + chunk <= nq) ? chunk : (nq - start);
+        struct ggml_tensor* t_qv = ggml_view_2d(ctx, t_query, qdim, count,
+                                                t_query->nb[1],
+                                                (size_t)start * t_query->nb[1]);
+        t_dots[g] = t_qv ? ggml_mul_mat(ctx, t_nodes, t_qv) : NULL;
+        if (!t_dots[g]) { built = false; break; }
+        ggml_build_forward_expand(gf, t_dots[g]);
+    }
+    if (!built) { free(t_dots); ggml_free(ctx); free(valid_orig); free(node_norm_sq); free(node_matrix); free(q_norm_sq); return false; }
+
+    struct ggml_backend_buffer* buf = ggml_backend_alloc_ctx_tensors(ctx, g_backend);
+    if (!buf) { free(t_dots); ggml_free(ctx); free(valid_orig); free(node_norm_sq); free(node_matrix); free(q_norm_sq); return false; }
+
+    ggml_backend_tensor_set(t_nodes, node_matrix, 0, (size_t)n_valid * qdim * sizeof(float));
+    ggml_backend_tensor_set(t_query, queries, 0, (size_t)nq * qdim * sizeof(float));
+    free(node_matrix); /* uploaded; the packed CPU copy is no longer needed */
+
+    enum ggml_status st = ggml_backend_graph_compute(g_backend, gf);
+    if (st != GGML_STATUS_SUCCESS) {
+        free(t_dots); ggml_backend_buffer_free(buf); ggml_free(ctx);
+        free(valid_orig); free(node_norm_sq); free(q_norm_sq);
+        return false;
+    }
+
+    /* Each group's result is [n_valid, count] contiguous, so reading group g
+     * into dot + start*n_valid reconstructs exactly the same flat
+     * dot[j*n_valid + w] layout a single ne11=nq mul_mat would have produced —
+     * Pass 3 below is unchanged by the chunking. */
+    float* dot = (float*)malloc((size_t)n_valid * (size_t)nq * sizeof(float));
+    if (!dot) { free(t_dots); ggml_backend_buffer_free(buf); ggml_free(ctx); free(valid_orig); free(node_norm_sq); free(q_norm_sq); return false; }
+    for (int32_t g = 0; g < ngroups; g++) {
+        const int32_t start = g * chunk;
+        const int32_t count = (start + chunk <= nq) ? chunk : (nq - start);
+        ggml_backend_tensor_get(t_dots[g], dot + (size_t)start * n_valid, 0,
+                                (size_t)count * (size_t)n_valid * sizeof(float));
+    }
+    free(t_dots);
+
+    /* Pass 3 (CPU): combine dot/(||a||*||b||) per (query,node) pair, scatter
+     * into out_scores at ORIGINAL node indices; every excluded row gets
+     * -2.0 for every query. out_scores is fully populated either way. */
+    for (int32_t j = 0; j < nq; j++) {
+        double* orow = out_scores + (size_t)j * n;
+        for (int32_t i = 0; i < n; i++) orow[i] = -2.0; /* default: excluded */
+        for (int32_t w = 0; w < n_valid; w++) {
+            float na = node_norm_sq[w], nb = q_norm_sq[j];
+            int32_t oi = valid_orig[w];
+            if (na <= 0.0f || nb <= 0.0f) { orow[oi] = -2.0; continue; }
+            float d = dot[(size_t)j * n_valid + w];
+            orow[oi] = (double)(d / sqrtf(na * nb));
+        }
+    }
+
+    free(dot);
+    ggml_backend_buffer_free(buf);
+    ggml_free(ctx);
+    free(valid_orig); free(node_norm_sq); free(q_norm_sq);
+    return true;
+}
+
+static bool ggml_strategy_batch(const float* query, int32_t qdim,
+                                 const float* const* node_ptrs, const int32_t* node_dims,
+                                 int32_t n, double* out_scores) {
+    if (!query || qdim <= 0 || !node_ptrs || !node_dims || n <= 0 || !out_scores) return false;
+    /* out_scores here is n doubles (nq=1); eg_ggml_run writes n*nq = n of
+     * them, laid out identically to the single-query contract. */
+    return eg_ggml_run(query, qdim, 1, node_ptrs, node_dims, n, out_scores);
+}
+
+static bool ggml_strategy_batch_multi(const float* queries, int32_t qdim, int32_t nq,
+                                       const float* const* node_ptrs, const int32_t* node_dims,
+                                       int32_t n, double* out_scores) {
+    return eg_ggml_run(queries, qdim, nq, node_ptrs, node_dims, n, out_scores);
+}
+
+static const EgCosineBatchStrategy g_ggml_strategy = {
+    .name = "ggml",
+    .available = ggml_strategy_available,
+    .batch = ggml_strategy_batch,
+    .batch_multi = ggml_strategy_batch_multi,
+};
+
+const EgCosineBatchStrategy* eg_cosine_batch_strategy_ggml(void) {
+    return &g_ggml_strategy;
+}

lang/runtime/eg_cosine_batch_strategy_metal_hand.m

@@ -0,0 +1,358 @@
+/* eg_cosine_batch_strategy_metal_hand.m — the HAND-ROLLED-METAL Strategy.
+ *
+ * This is PR #114's original Objective-C bridge (formerly eg_metal_cosine.m)
+ * exposing the hand-written Metal compute shader (eg_cosine_batch.metal) as
+ * one concrete EgCosineBatchStrategy. It is preserved here almost verbatim —
+ * real, carefully verified work, not discarded — now living behind the
+ * Adapter/Strategy/Factory restructuring (see eg_cosine_batch.h and
+ * eg_cosine_batch_strategy.h) alongside the new ggml-Metal strategy
+ * (eg_cosine_batch_strategy_ggml.c) and the universal CPU fallback
+ * (eg_cosine_batch_strategy_cpu.c). The factory in eg_cosine_batch.c prefers
+ * ggml by default when both are available; this strategy remains selectable
+ * via EL_COSINE_BATCH_STRATEGY=metal, and is what the factory falls back to
+ * if ggml's backend plugin fails to load/init for any reason.
+ *
+ * Apple-only (Metal has no other platform). This file is excluded from the
+ * build entirely on non-Darwin — see build_vindex_bench.sh, which only
+ * compiles/links this file and defines EG_HAVE_STRATEGY_METAL_HAND when
+ * `uname` is Darwin. On Linux the factory never sees this strategy at all —
+ * callers must always be prepared for the "no real strategy available"
+ * fallback via the CPU strategy, which is also exactly what happens here on
+ * Apple hardware with no usable GPU.
+ *
+ * Design (unchanged from PR #114):
+ *   - Device/queue/pipeline are created lazily, once, and cached in static
+ *     globals — every call after the first only allocates buffers + submits.
+ *   - The Metal shader source is embedded as a C string literal (kMetalSrc
+ *     below) rather than loaded from a file at runtime or shipped as a
+ *     precompiled .metallib. Chosen over newLibraryWithFile: /a .metallib
+ *     because the engram binary can be invoked from an arbitrary working
+ *     directory (launchd job, nsbx sandbox, CI) and a file-path shader would
+ *     be one relocation away from silently falling back to CPU for reasons
+ *     that have nothing to do with Metal availability. Embedding costs one
+ *     runtime shader compile (~tens of ms) on first use, amortized over the
+ *     process lifetime, in exchange for a genuinely self-contained binary.
+ *     Source of truth for review/tooling is eg_cosine_batch.metal — this
+ *     string MUST be kept byte-identical to that file (a comment marks both
+ *     ends of the copy).
+ *   - Buffers use MTLResourceStorageModeShared: on Apple Silicon's unified
+ *     memory, CPU and GPU read the same physical pages, so filling a buffer
+ *     is a plain memcpy and there is no separate "upload" step.
+ *   - ANY failure at ANY step (no device, pipeline compile error, buffer
+ *     allocation failure, bad args) returns false and leaves out_scores
+ *     untouched. This function is called from the request-handling hot path
+ *     of a long-lived daemon — it must never throw, crash, or hang it.
+ */
+#import <Foundation/Foundation.h>
+#import <Metal/Metal.h>
+#include "eg_cosine_batch_strategy.h"
+#include <string.h>
+#include <stdlib.h>
+
+/* ── BEGIN embedded shader source (keep in sync with eg_cosine_batch.metal) ── */
+static const char* kEgCosineBatchMetalSrc =
+"#include <metal_stdlib>\n"
+"using namespace metal;\n"
+"struct EgCosineParams { uint n; uint dim; };\n"
+"kernel void eg_cosine_batch_kernel(\n"
+"    device const float*  query        [[buffer(0)]],\n"
+"    device const float*  node_matrix  [[buffer(1)]],\n"
+"    device const int*    node_dims    [[buffer(2)]],\n"
+"    constant EgCosineParams& p        [[buffer(3)]],\n"
+"    device float*        out_scores   [[buffer(4)]],\n"
+"    uint gid [[thread_position_in_grid]])\n"
+"{\n"
+"    if (gid >= p.n) return;\n"
+"    if (node_dims[gid] != int(p.dim)) { out_scores[gid] = -2.0f; return; }\n"
+"    device const float* row = node_matrix + (uint64_t)gid * (uint64_t)p.dim;\n"
+"    float dot0 = 0.0f, dot1 = 0.0f, dot2 = 0.0f, dot3 = 0.0f;\n"
+"    float na0  = 0.0f, na1  = 0.0f, na2  = 0.0f, na3  = 0.0f;\n"
+"    float nb0  = 0.0f, nb1  = 0.0f, nb2  = 0.0f, nb3  = 0.0f;\n"
+"    uint d = 0;\n"
+"    uint dim4 = p.dim & ~3u;\n"
+"    for (; d < dim4; d += 4) {\n"
+"        float a0 = row[d],   b0 = query[d];\n"
+"        float a1 = row[d+1], b1 = query[d+1];\n"
+"        float a2 = row[d+2], b2 = query[d+2];\n"
+"        float a3 = row[d+3], b3 = query[d+3];\n"
+"        dot0 += a0*b0; dot1 += a1*b1; dot2 += a2*b2; dot3 += a3*b3;\n"
+"        na0  += a0*a0; na1  += a1*a1; na2  += a2*a2; na3  += a3*a3;\n"
+"        nb0  += b0*b0; nb1  += b1*b1; nb2  += b2*b2; nb3  += b3*b3;\n"
+"    }\n"
+"    float dot = (dot0 + dot1) + (dot2 + dot3);\n"
+"    float na  = (na0  + na1)  + (na2  + na3);\n"
+"    float nb  = (nb0  + nb1)  + (nb2  + nb3);\n"
+"    for (; d < p.dim; d++) {\n"
+"        float a = row[d], b = query[d];\n"
+"        dot += a*b; na += a*a; nb += b*b;\n"
+"    }\n"
+"    if (na <= 0.0f || nb <= 0.0f) { out_scores[gid] = -2.0f; return; }\n"
+"    out_scores[gid] = dot / sqrt(na * nb);\n"
+"}\n"
+"struct EgCosineMultiParams { uint n; uint dim; uint nq; };\n"
+"kernel void eg_cosine_batch_multi_kernel(\n"
+"    device const float*  queries      [[buffer(0)]],\n"
+"    device const float*  node_matrix  [[buffer(1)]],\n"
+"    device const int*    node_dims    [[buffer(2)]],\n"
+"    constant EgCosineMultiParams& p   [[buffer(3)]],\n"
+"    device float*        out_scores   [[buffer(4)]],\n"
+"    uint2 gid [[thread_position_in_grid]])\n"
+"{\n"
+"    uint nid = gid.x, qid = gid.y;\n"
+"    if (nid >= p.n || qid >= p.nq) return;\n"
+"    uint64_t out_idx = (uint64_t)qid * (uint64_t)p.n + (uint64_t)nid;\n"
+"    if (node_dims[nid] != int(p.dim)) { out_scores[out_idx] = -2.0f; return; }\n"
+"    device const float* row   = node_matrix + (uint64_t)nid * (uint64_t)p.dim;\n"
+"    device const float* query = queries     + (uint64_t)qid * (uint64_t)p.dim;\n"
+"    float dot0 = 0.0f, dot1 = 0.0f, dot2 = 0.0f, dot3 = 0.0f;\n"
+"    float na0  = 0.0f, na1  = 0.0f, na2  = 0.0f, na3  = 0.0f;\n"
+"    float nb0  = 0.0f, nb1  = 0.0f, nb2  = 0.0f, nb3  = 0.0f;\n"
+"    uint d = 0;\n"
+"    uint dim4 = p.dim & ~3u;\n"
+"    for (; d < dim4; d += 4) {\n"
+"        float a0 = row[d],   b0 = query[d];\n"
+"        float a1 = row[d+1], b1 = query[d+1];\n"
+"        float a2 = row[d+2], b2 = query[d+2];\n"
+"        float a3 = row[d+3], b3 = query[d+3];\n"
+"        dot0 += a0*b0; dot1 += a1*b1; dot2 += a2*b2; dot3 += a3*b3;\n"
+"        na0  += a0*a0; na1  += a1*a1; na2  += a2*a2; na3  += a3*a3;\n"
+"        nb0  += b0*b0; nb1  += b1*b1; nb2  += b2*b2; nb3  += b3*b3;\n"
+"    }\n"
+"    float dot = (dot0 + dot1) + (dot2 + dot3);\n"
+"    float na  = (na0  + na1)  + (na2  + na3);\n"
+"    float nb  = (nb0  + nb1)  + (nb2  + nb3);\n"
+"    for (; d < p.dim; d++) {\n"
+"        float a = row[d], b = query[d];\n"
+"        dot += a*b; na += a*a; nb += b*b;\n"
+"    }\n"
+"    if (na <= 0.0f || nb <= 0.0f) { out_scores[out_idx] = -2.0f; return; }\n"
+"    out_scores[out_idx] = dot / sqrt(na * nb);\n"
+"}\n";
+/* ── END embedded shader source ── */
+
+typedef struct EgCosineParamsC { uint32_t n; uint32_t dim; } EgCosineParamsC;
+typedef struct EgCosineMultiParamsC { uint32_t n; uint32_t dim; uint32_t nq; } EgCosineMultiParamsC;
+
+static id<MTLDevice>              g_device   = nil;
+static id<MTLCommandQueue>        g_queue    = nil;
+static id<MTLComputePipelineState> g_pipeline = nil;       /* single-query kernel */
+static id<MTLComputePipelineState> g_pipeline_multi = nil; /* multi-query kernel */
+static bool g_init_attempted = false;
+static bool g_init_ok        = false;
+
+/* Lazy, one-time setup. Never throws — every Metal call here is the
+ * "returns nil/NSError on failure" flavor, not an exception-throwing one. */
+static bool eg_metal_ensure_init(void) {
+    if (g_init_attempted) return g_init_ok;
+    g_init_attempted = true;
+
+    @autoreleasepool {
+        id<MTLDevice> dev = MTLCreateSystemDefaultDevice();
+        if (!dev) return false;
+
+        id<MTLCommandQueue> q = [dev newCommandQueue];
+        if (!q) return false;
+
+        NSError* err = nil;
+        NSString* src = [NSString stringWithUTF8String:kEgCosineBatchMetalSrc];
+        MTLCompileOptions* opts = [MTLCompileOptions new];
+        id<MTLLibrary> lib = [dev newLibraryWithSource:src options:opts error:&err];
+        if (!lib) return false;
+
+        id<MTLFunction> fn = [lib newFunctionWithName:@"eg_cosine_batch_kernel"];
+        if (!fn) return false;
+        id<MTLComputePipelineState> pipe = [dev newComputePipelineStateWithFunction:fn error:&err];
+        if (!pipe) return false;
+
+        id<MTLFunction> fnMulti = [lib newFunctionWithName:@"eg_cosine_batch_multi_kernel"];
+        if (!fnMulti) return false;
+        id<MTLComputePipelineState> pipeMulti = [dev newComputePipelineStateWithFunction:fnMulti error:&err];
+        if (!pipeMulti) return false;
+
+        g_device = dev;
+        g_queue = q;
+        g_pipeline = pipe;
+        g_pipeline_multi = pipeMulti;
+        g_init_ok = true;
+        return true;
+    }
+}
+
+static bool mh_available(void) {
+    return eg_metal_ensure_init();
+}
+
+static bool mh_batch(const float* query, int32_t qdim,
+                      const float* const* node_ptrs,
+                      const int32_t* node_dims,
+                      int32_t n,
+                      double* out_scores) {
+    if (!query || qdim <= 0 || !node_ptrs || !node_dims || n <= 0 || !out_scores) return false;
+    if (!eg_metal_ensure_init()) return false;
+
+    @autoreleasepool {
+        const size_t dim = (size_t)qdim;
+        const size_t nu  = (size_t)n;
+
+        /* Gather into a packed row-major matrix — EngramNode.emb is one
+         * malloc per node, not a contiguous array, so this copy is
+         * unavoidable regardless of backend. Rows whose real dim doesn't
+         * match qdim are zero-filled (harmless: the kernel sentinels them
+         * via node_dims before ever reading the row). */
+        float* matrix = (float*)calloc(nu * dim, sizeof(float));
+        int32_t* dims_i32 = (int32_t*)malloc(nu * sizeof(int32_t));
+        if (!matrix || !dims_i32) { free(matrix); free(dims_i32); return false; }
+
+        for (size_t i = 0; i < nu; i++) {
+            dims_i32[i] = node_dims[i];
+            if (node_ptrs[i] && node_dims[i] == qdim) {
+                memcpy(matrix + i * dim, node_ptrs[i], dim * sizeof(float));
+            }
+            /* else: leave zero-filled; node_dims[i] != qdim (or missing)
+             * makes the kernel sentinel it to -2.0 without reading the row. */
+        }
+
+        id<MTLBuffer> bufQuery = [g_device newBufferWithBytes:query
+                                                        length:dim * sizeof(float)
+                                                       options:MTLResourceStorageModeShared];
+        id<MTLBuffer> bufMatrix = [g_device newBufferWithBytes:matrix
+                                                         length:nu * dim * sizeof(float)
+                                                        options:MTLResourceStorageModeShared];
+        id<MTLBuffer> bufDims = [g_device newBufferWithBytes:dims_i32
+                                                       length:nu * sizeof(int32_t)
+                                                      options:MTLResourceStorageModeShared];
+        EgCosineParamsC params = { (uint32_t)nu, (uint32_t)dim };
+        id<MTLBuffer> bufParams = [g_device newBufferWithBytes:&params
+                                                         length:sizeof(params)
+                                                        options:MTLResourceStorageModeShared];
+        id<MTLBuffer> bufOut = [g_device newBufferWithLength:nu * sizeof(float)
+                                                       options:MTLResourceStorageModeShared];
+
+        free(matrix); free(dims_i32);
+
+        if (!bufQuery || !bufMatrix || !bufDims || !bufParams || !bufOut) return false;
+
+        id<MTLCommandBuffer> cmd = [g_queue commandBuffer];
+        if (!cmd) return false;
+        id<MTLComputeCommandEncoder> enc = [cmd computeCommandEncoder];
+        if (!enc) return false;
+
+        [enc setComputePipelineState:g_pipeline];
+        [enc setBuffer:bufQuery  offset:0 atIndex:0];
+        [enc setBuffer:bufMatrix offset:0 atIndex:1];
+        [enc setBuffer:bufDims   offset:0 atIndex:2];
+        [enc setBuffer:bufParams offset:0 atIndex:3];
+        [enc setBuffer:bufOut    offset:0 atIndex:4];
+
+        NSUInteger tgSize = g_pipeline.maxTotalThreadsPerThreadgroup;
+        if (tgSize > 256) tgSize = 256;
+        if (tgSize < 1) tgSize = 1;
+        MTLSize gridSize = MTLSizeMake(nu, 1, 1);
+        MTLSize threadgroupSize = MTLSizeMake(tgSize, 1, 1);
+        [enc dispatchThreads:gridSize threadsPerThreadgroup:threadgroupSize];
+        [enc endEncoding];
+
+        [cmd commit];
+        [cmd waitUntilCompleted];
+
+        if (cmd.status != MTLCommandBufferStatusCompleted) return false;
+
+        const float* results = (const float*)bufOut.contents;
+        if (!results) return false;
+        for (size_t i = 0; i < nu; i++) out_scores[i] = (double)results[i];
+        return true;
+    }
+}
+
+static bool mh_batch_multi(const float* queries, int32_t qdim, int32_t nq,
+                            const float* const* node_ptrs,
+                            const int32_t* node_dims,
+                            int32_t n,
+                            double* out_scores) {
+    if (!queries || qdim <= 0 || nq <= 0 || !node_ptrs || !node_dims || n <= 0 || !out_scores) return false;
+    if (!eg_metal_ensure_init()) return false;
+
+    @autoreleasepool {
+        const size_t dim = (size_t)qdim;
+        const size_t nu  = (size_t)n;
+        const size_t nqu = (size_t)nq;
+
+        float* matrix = (float*)calloc(nu * dim, sizeof(float));
+        int32_t* dims_i32 = (int32_t*)malloc(nu * sizeof(int32_t));
+        if (!matrix || !dims_i32) { free(matrix); free(dims_i32); return false; }
+
+        for (size_t i = 0; i < nu; i++) {
+            dims_i32[i] = node_dims[i];
+            if (node_ptrs[i] && node_dims[i] == qdim) {
+                memcpy(matrix + i * dim, node_ptrs[i], dim * sizeof(float));
+            }
+        }
+
+        /* This is the ONE upload of node_matrix for the whole nq-query batch —
+         * the fix for the measured re-upload-per-query slowdown. */
+        id<MTLBuffer> bufMatrix = [g_device newBufferWithBytes:matrix
+                                                         length:nu * dim * sizeof(float)
+                                                        options:MTLResourceStorageModeShared];
+        id<MTLBuffer> bufDims = [g_device newBufferWithBytes:dims_i32
+                                                       length:nu * sizeof(int32_t)
+                                                      options:MTLResourceStorageModeShared];
+        id<MTLBuffer> bufQueries = [g_device newBufferWithBytes:queries
+                                                          length:nqu * dim * sizeof(float)
+                                                         options:MTLResourceStorageModeShared];
+        EgCosineMultiParamsC params = { (uint32_t)nu, (uint32_t)dim, (uint32_t)nqu };
+        id<MTLBuffer> bufParams = [g_device newBufferWithBytes:&params
+                                                         length:sizeof(params)
+                                                        options:MTLResourceStorageModeShared];
+        id<MTLBuffer> bufOut = [g_device newBufferWithLength:nqu * nu * sizeof(float)
+                                                       options:MTLResourceStorageModeShared];
+
+        free(matrix); free(dims_i32);
+
+        if (!bufMatrix || !bufDims || !bufQueries || !bufParams || !bufOut) return false;
+
+        id<MTLCommandBuffer> cmd = [g_queue commandBuffer];
+        if (!cmd) return false;
+        id<MTLComputeCommandEncoder> enc = [cmd computeCommandEncoder];
+        if (!enc) return false;
+
+        [enc setComputePipelineState:g_pipeline_multi];
+        [enc setBuffer:bufQueries offset:0 atIndex:0];
+        [enc setBuffer:bufMatrix  offset:0 atIndex:1];
+        [enc setBuffer:bufDims    offset:0 atIndex:2];
+        [enc setBuffer:bufParams  offset:0 atIndex:3];
+        [enc setBuffer:bufOut     offset:0 atIndex:4];
+
+        /* 2D dispatch: x over nodes, y over queries. Threadgroup width picked
+         * from the pipeline's own limit, height fixed at 1 — nq is typically
+         * small (tens to low hundreds) relative to n (thousands+), so tiling
+         * the wide axis (n) is what matters for occupancy. */
+        NSUInteger tgWidth = g_pipeline_multi.maxTotalThreadsPerThreadgroup;
+        if (tgWidth > 256) tgWidth = 256;
+        if (tgWidth < 1) tgWidth = 1;
+        MTLSize gridSize = MTLSizeMake(nu, nqu, 1);
+        MTLSize threadgroupSize = MTLSizeMake(tgWidth, 1, 1);
+        [enc dispatchThreads:gridSize threadsPerThreadgroup:threadgroupSize];
+        [enc endEncoding];
+
+        [cmd commit];
+        [cmd waitUntilCompleted];
+
+        if (cmd.status != MTLCommandBufferStatusCompleted) return false;
+
+        const float* results = (const float*)bufOut.contents;
+        if (!results) return false;
+        for (size_t i = 0; i < nqu * nu; i++) out_scores[i] = (double)results[i];
+        return true;
+    }
+}
+
+static const EgCosineBatchStrategy g_metal_hand_strategy = {
+    .name = "metal-hand",
+    .available = mh_available,
+    .batch = mh_batch,
+    .batch_multi = mh_batch_multi,
+};
+
+const EgCosineBatchStrategy* eg_cosine_batch_strategy_metal_hand(void) {
+    return &g_metal_hand_strategy;
+}

lang/runtime/el_runtime.h

@@ -275,6 +275,10 @@ el_val_t  json_set(el_val_t json_str, el_val_t key, el_val_t value);
 el_val_t  json_array_len(el_val_t json_str);
 el_val_t  json_array_get(el_val_t json_str, el_val_t index);
 el_val_t  json_array_get_string(el_val_t json_str, el_val_t index);
+el_val_t  json_escape_string(el_val_t sv);
+el_val_t  json_build_object(el_val_t kvs);
+el_val_t  json_build_array(el_val_t items);
+el_val_t  json_array_push(el_val_t arr_v, el_val_t elem_v);  /* defined in el_runtime.c */
 
 /* ── Time ────────────────────────────────────────────────────────────────── */
 
@@ -301,6 +305,8 @@ el_val_t  time_diff(el_val_t ts1, el_val_t ts2, el_val_t unit);
 
 el_val_t  el_now_instant(void);
 el_val_t  now(void);
+el_val_t  now_millis(void);   /* wall-clock milliseconds (defined in el_runtime.c) */
+el_val_t  now_ns(void);       /* wall-clock nanoseconds (defined in el_runtime.c) */
 el_val_t  unix_seconds(el_val_t n);
 el_val_t  unix_millis(el_val_t n);
 el_val_t  instant_from_iso8601(el_val_t s);
@@ -580,6 +586,60 @@ void      el_runtime_dharma_event_arrive(const char* event_type,
                                          const char* payload,
                                          const char* source);
 
+/* ── Geometry: signal as a first-class El value ──────────────────────────────
+ *
+ * A Geometry is an opaque, magic-tagged heap value carried in an el_val_t —
+ * the same discipline as List/Map. It holds a width and a float32 payload,
+ * and it is the medium a non-text modality enters in. Declared HERE, above
+ * the engram block, because transduction is a LANGUAGE concern: every El
+ * program touching any modality needs it, and the engram is merely one El
+ * program that happens to hold a graph. See el_runtime.c ("Geometry: signal
+ * as a first-class el value") for the full rationale.
+ *
+ * El-side type annotation is simply `Geometry` — an opaque boxed pointer,
+ * exactly like Instant / Calendar / Rhythm. No codegen change is required.
+ *
+ * OWNERSHIP: a Geometry is owned by the El caller and released with
+ * geometry_free. node_attach_geometry COPIES, so a node and the caller's
+ * value have independent lifetimes.                                          */
+
+el_val_t  geometry_new(el_val_t dim);                 /* zero-filled; 0 on failure */
+el_val_t  geometry_dim(el_val_t g);                   /* width, 0 if not a Geometry */
+el_val_t  geometry_is(el_val_t g);                    /* 1 if a live Geometry       */
+el_val_t  geometry_get(el_val_t g, el_val_t i);       /* Float component            */
+el_val_t  geometry_set(el_val_t g, el_val_t i, el_val_t x);  /* 1 ok / 0 out of range */
+el_val_t  geometry_norm(el_val_t g);                  /* Float L2 — lets a caller
+                                                       * check a realizer emitted
+                                                       * signal, not zeros          */
+el_val_t  geometry_free(el_val_t g);                  /* 1 if freed, 0 if not a Geometry.
+                                                       * Returns a value (not void) so it
+                                                       * is safe in any El expression
+                                                       * position without a codegen
+                                                       * void-builtin table entry.  */
+
+/* Wire ADAPTERS — the only place an encoding appears, and only at the edge.
+ * `f32le hex` is little-endian float32, 8 hex chars per component: the
+ * encoding the perception vessel's /voice/embed already emits. The width is
+ * DERIVED from the input length, never supplied by a caller — which is why
+ * there is no max-dim constant here to validate a claimed length against. */
+el_val_t  geometry_from_f32le_hex(el_val_t hex);      /* 0 on empty/odd-length/non-hex */
+el_val_t  geometry_to_f32le_hex(el_val_t g);          /* "" if not a Geometry       */
+
+/* ── Realizers + transduce ───────────────────────────────────────────────────
+ * A REALIZER maps one modality into geometry. Registration is by NAME, so a
+ * new modality never requires a runtime patch: every El `fn name(...)`
+ * compiles to a global C symbol with that exact name, and the registry
+ * resolves it with dlsym against the running binary — the same mechanism
+ * http_set_handler already relies on.
+ *
+ *     fn tone_realizer(signal: String) -> Geometry { ... }
+ *     realizer_register("tone", "tone_realizer")
+ *     let g: Geometry = transduce(sample, "tone")
+ */
+el_val_t  realizer_register(el_val_t modality, el_val_t fn_name); /* 1 ok / 0 unresolved */
+el_val_t  realizer_has(el_val_t modality);            /* 1 if a realizer is registered */
+el_val_t  transduce(el_val_t signal, el_val_t modality); /* Geometry, or 0 if no organ */
+
 /* ── Engram local graph primitives ───────────────────────────────────────────
  * Operate on the CGI's local Engram knowledge graph.
  * `engram_activate` queries the local graph only; `dharma_activate` is
@@ -607,6 +667,23 @@ void      engram_strengthen(el_val_t node_id);
 void      engram_forget(el_val_t node_id);
 el_val_t  engram_prune_telemetry(el_val_t older_than_ms);
 el_val_t  engram_node_count(void);
+/* Attach a Geometry to an existing node, and read the attached width back.
+ * Named for the operation, not the store: a node acquires geometry. This is
+ * the geometry-valued ingest path — nothing about it is hex, and nothing
+ * about it assumes the caller's vector matches the canonical text-embedding
+ * width. node_geometry_dim exists so an attach is VERIFIED by reading it
+ * back rather than by trusting a success return. */
+el_val_t  node_attach_geometry(el_val_t node_id, el_val_t g); /* 1 ok / 0 otherwise */
+el_val_t  node_geometry_dim(el_val_t node_id);                /* width, 0 if none   */
+
+/* DEPRECATED (shipped in #141, superseded 2026-08-16). Equivalent to
+ * geometry_from_f32le_hex + node_attach_geometry, and now implemented as
+ * exactly that. Kept only so anything built against the #141 runtime keeps
+ * linking; `dim` is accepted but treated as an assertion about the vector's
+ * width rather than as its source. New code should not call this — a hex
+ * string is a wire encoding, not a way to move geometry between two pieces
+ * of El. Returns 1 on success, 0 otherwise. */
+el_val_t  engram_node_set_emb(el_val_t id, el_val_t hex, el_val_t dim);
 el_val_t  engram_search(el_val_t query, el_val_t limit);
 el_val_t  engram_scan_nodes(el_val_t limit, el_val_t offset);
 void      engram_connect(el_val_t from_id, el_val_t to_id, el_val_t weight, el_val_t relation);
@@ -712,6 +789,14 @@ el_val_t  engram_label_df(el_val_t term);
 el_val_t  engram_salient_term(el_val_t node_id, el_val_t max_df,
                               el_val_t min_df, el_val_t tabu);
 el_val_t  engram_embed_backfill(el_val_t count);
+/* op_assert seam: grounded assertion envelope {subject,grounding} for the realizer. */
+el_val_t  engram_op_assert_json(el_val_t node_id, el_val_t depth);
+/* Parametric mutation (purview write-side): purview==0 => G=live (default), else refuse. */
+el_val_t  engram_node_full_in(el_val_t purview, el_val_t content, el_val_t node_type, el_val_t label,
+                              el_val_t salience, el_val_t importance, el_val_t confidence,
+                              el_val_t tier, el_val_t tags);
+void      engram_connect_in(el_val_t purview, el_val_t from_id, el_val_t to_id,
+                            el_val_t weight, el_val_t relation);
 el_val_t  engram_list_layers_json(void);
 /* Working memory introspection — count, mean weight, and top-N snapshot.
  * Ported from runtime on 2026-06-30 self-review. */
@@ -892,6 +977,140 @@ el_val_t trace_span_start(el_val_t name);
 el_val_t trace_span_end(el_val_t span_handle);
 el_val_t emit_event(el_val_t name, el_val_t duration_ms);
 
+el_val_t __thread_create(el_val_t fn_name_v, el_val_t arg_v);
+el_val_t __thread_join(el_val_t tid_v);
+
+/* Mutex + channel seed primitives (defined in el_runtime.c). Declared here so
+ * that compiled El programs which use runtime/thread.el's with_mutex helper or
+ * runtime/channel.el's Go-style channels see real prototypes instead of an
+ * implicit int-return declaration (which the C11 ABI mis-truncates el_val_t). */
+el_val_t __mutex_new(void);
+void     __mutex_lock(el_val_t m_v);
+void     __mutex_unlock(el_val_t m_v);
+el_val_t __channel_new(el_val_t capacity_v);
+el_val_t __channel_send(el_val_t ch_v, el_val_t msg_v);
+el_val_t __channel_recv(el_val_t ch_v);
+el_val_t __channel_try_recv(el_val_t ch_v);
+el_val_t __channel_close(el_val_t ch_v);
+
+/* ── __ prefixed aliases (self-hosting compiler ABI) ─────────────────────────
+ * The El self-hosting compiler emits calls to __-prefixed names. These are
+ * forwarding wrappers around the existing el_runtime functions above.        */
+
+/* I/O */
+el_val_t  __println(el_val_t s);
+el_val_t  __print(el_val_t s);
+el_val_t  __readline(void);
+
+/* String */
+el_val_t  __int_to_str(el_val_t n);
+el_val_t  __str_to_int(el_val_t s);
+el_val_t  __float_to_str(el_val_t f);
+el_val_t  __str_to_float(el_val_t s);
+el_val_t  __str_len(el_val_t s);
+el_val_t  __str_char_at(el_val_t s, el_val_t i);
+el_val_t  __str_cmp(el_val_t a, el_val_t b);
+el_val_t  __str_ncmp(el_val_t a, el_val_t b, el_val_t n);
+el_val_t  __str_concat_raw(el_val_t a, el_val_t b);
+el_val_t  __str_slice_raw(el_val_t s, el_val_t start, el_val_t end);
+el_val_t  __str_alloc(el_val_t n);
+el_val_t  __str_set_char(el_val_t s, el_val_t i, el_val_t c);
+
+/* URL encoding */
+el_val_t  __url_encode(el_val_t s);
+el_val_t  __url_decode(el_val_t s);
+
+/* Environment */
+el_val_t  __env_get(el_val_t key);
+
+/* Subprocess */
+el_val_t  __exec(el_val_t cmd);
+el_val_t  __exec_bg(el_val_t cmd);
+
+/* Process */
+el_val_t  __exit_program(el_val_t code);
+
+/* Filesystem */
+el_val_t  __fs_exists(el_val_t path);
+el_val_t  __fs_mkdir(el_val_t path);
+el_val_t  __fs_read(el_val_t path);
+el_val_t  __fs_write(el_val_t path, el_val_t content);
+el_val_t  __fs_write_bytes(el_val_t path, el_val_t bytes, el_val_t n);
+el_val_t  __fs_list_raw(el_val_t path);
+
+/* HTTP server */
+el_val_t  __http_response(el_val_t status, el_val_t headers_json, el_val_t body);
+el_val_t  __http_serve(el_val_t port, el_val_t handler);
+el_val_t  __http_serve_v2(el_val_t port, el_val_t handler);
+
+/* HTTP conn fd / SSE (weak; overridden by el_seed.c when linked together) */
+el_val_t  __http_conn_fd(void);
+el_val_t  __http_sse_open(el_val_t conn_id);
+el_val_t  __http_sse_send(el_val_t conn_id, el_val_t data);
+el_val_t  __http_sse_close(el_val_t conn_id);
+
+/* HTTP client (requires HAVE_CURL; stubs provided for no-curl builds) */
+el_val_t  __http_do(el_val_t method, el_val_t url, el_val_t body,
+                    el_val_t headers_map, el_val_t timeout_ms);
+el_val_t  __http_do_map(el_val_t method, el_val_t url, el_val_t body,
+                        el_val_t headers_json, el_val_t timeout_ms);
+el_val_t  __http_do_map_to_file(el_val_t method, el_val_t url, el_val_t body,
+                                el_val_t headers_json, el_val_t output_path);
+
+/* JSON */
+el_val_t  __json_array_get(el_val_t json, el_val_t index);
+el_val_t  __json_array_get_string(el_val_t json, el_val_t index);
+el_val_t  __json_array_len(el_val_t json);
+el_val_t  __json_get(el_val_t json, el_val_t key);
+el_val_t  __json_get_raw(el_val_t json, el_val_t key);
+el_val_t  __json_set(el_val_t json, el_val_t key, el_val_t value);
+el_val_t  __json_parse_map(el_val_t json_str);
+el_val_t  __json_stringify_val(el_val_t val);
+
+/* Hashing */
+el_val_t  __sha256_hex(el_val_t s);
+
+/* State K/V */
+el_val_t  __state_del(el_val_t key);
+el_val_t  __state_get(el_val_t key);
+el_val_t  __state_keys(void);
+el_val_t  __state_set(el_val_t key, el_val_t val);
+
+/* UUID */
+el_val_t  __uuid_v4(void);
+
+/* Args */
+el_val_t  __args_json(void);
+
+/* Compiler-support builtins — called by the El compiler's own source
+ * (compiler.el, codegen.el) and registered in codegen.el's builtin_arity. */
+el_val_t  stdout_to_file(el_val_t path);
+el_val_t  stdout_restore(void);
+el_val_t  el_mem_check(void);
+
+/* Allocation accounting — the deterministic signal behind complexity gating.
+ * Gate on counts/bytes; peak RSS is context only. */
+el_val_t  el_alloc_count(void);
+el_val_t  el_alloc_bytes(void);
+el_val_t  el_peak_rss(void);
+el_val_t  el_black_box(el_val_t v);
+
+/* Semantic retrieval surface. NOT interchangeable with engram_search_json,
+ * which is lexical by design — see the note at the definition. */
+el_val_t  engram_recall_json(el_val_t query, el_val_t limit);
+
+/* Edges straight from the store — replaces the engram_save()+fs_read()
+ * whole-graph round trip that /api/graph/edges used to do. */
+el_val_t  engram_edges_json(el_val_t limit, el_val_t offset);
+
+/* Buffer-pool interoception as JSON — live pool health for observation. */
+el_val_t  engram_pool_stats_json(void);
+
+/* CGI identity accessors (read-only). */
+el_val_t  cgi_principal(void);
+el_val_t  cgi_network(void);
+el_val_t  cgi_engram(void);
+
 #ifdef __cplusplus
 }
 #endif

lang/runtime/el_seed.c

@@ -37,6 +37,82 @@
 #include <dlfcn.h>
 #include <curl/curl.h>
 
+/* el_runtime.c bridge prototypes.
+ *
+ * A block of __-prefixed wrappers further down in this file (http serving,
+ * JSON access, key-val state, URL/HTML escaping, and the whole engram_*
+ * node/edge/layer/search surface -- 51 symbols in total) delegate to
+ * unprefixed counterparts that are implemented in el_runtime.c, not here.
+ * Porting them into native el_seed.c or El has not happened yet.
+ * tools/install.sh compiles el_seed.c and el_runtime.c as separate objects
+ * and archives both into libel.a, so the symbols are always present at link
+ * time. el_seed.c alone was just missing the prototypes, which made even a
+ * standalone -c compile of this one file fail on a toolchain that now treats
+ * an implicit function declaration as a hard error under C11.
+ *
+ * A plain include of el_runtime.h was tried first and rejected: it redefines
+ * el_to_float and el_from_float, which el_seed.h already provides. Narrow
+ * prototypes, copied verbatim from el_runtime.h, avoid that collision without
+ * pulling in the rest of the retiring runtime header.
+ */
+el_val_t http_response(el_val_t status, el_val_t headers_json, el_val_t body);
+void     http_serve(el_val_t port, el_val_t handler);
+void     http_serve_v2(el_val_t port, el_val_t handler);
+el_val_t json_get(el_val_t json, el_val_t key);
+el_val_t json_get_string(el_val_t json_str, el_val_t key);
+el_val_t json_get_int(el_val_t json_str, el_val_t key);
+el_val_t json_get_float(el_val_t json_str, el_val_t key);
+el_val_t json_get_bool(el_val_t json_str, el_val_t key);
+el_val_t json_get_raw(el_val_t json_str, el_val_t key);
+el_val_t json_parse(el_val_t s);
+el_val_t json_set(el_val_t json_str, el_val_t key, el_val_t value);
+el_val_t json_stringify(el_val_t v);
+el_val_t json_array_len(el_val_t json_str);
+el_val_t json_array_get(el_val_t json_str, el_val_t index);
+el_val_t json_array_get_string(el_val_t json_str, el_val_t index);
+el_val_t state_set(el_val_t key, el_val_t value);
+el_val_t state_get(el_val_t key);
+el_val_t state_del(el_val_t key);
+el_val_t state_keys(void);
+el_val_t url_encode(el_val_t s);
+el_val_t url_decode(el_val_t s);
+el_val_t el_html_sanitize(el_val_t input_html, el_val_t allowlist_json);
+el_val_t engram_node(el_val_t content, el_val_t node_type, el_val_t salience);
+el_val_t engram_node_full(el_val_t content, el_val_t node_type, el_val_t label,
+                           el_val_t salience, el_val_t importance, el_val_t confidence,
+                           el_val_t tier, el_val_t tags);
+el_val_t engram_node_layered(el_val_t content, el_val_t node_type, el_val_t label,
+                              el_val_t salience, el_val_t certainty, el_val_t confidence,
+                              el_val_t status, el_val_t tags, el_val_t layer_id);
+el_val_t engram_add_layer(el_val_t name, el_val_t priority, el_val_t suppressible,
+                           el_val_t transparent, el_val_t injectable);
+el_val_t engram_remove_layer(el_val_t layer_id);
+el_val_t engram_list_layers(void);
+el_val_t engram_list_layers_json(void);
+el_val_t engram_get_node(el_val_t id);
+el_val_t engram_get_node_json(el_val_t id);
+el_val_t engram_get_node_by_label(el_val_t label);
+void     engram_strengthen(el_val_t node_id);
+void     engram_forget(el_val_t node_id);
+el_val_t engram_node_count(void);
+el_val_t engram_edge_count(void);
+el_val_t engram_scan_nodes(el_val_t limit, el_val_t offset);
+el_val_t engram_scan_nodes_json(el_val_t limit, el_val_t offset);
+el_val_t engram_scan_nodes_by_type_json(el_val_t node_type, el_val_t limit, el_val_t offset);
+el_val_t engram_search(el_val_t query, el_val_t limit);
+el_val_t engram_search_json(el_val_t query, el_val_t limit);
+el_val_t engram_activate(el_val_t query, el_val_t depth);
+el_val_t engram_activate_json(el_val_t query, el_val_t depth);
+el_val_t engram_compile_layered_json(el_val_t intent, el_val_t depth);
+el_val_t engram_stats_json(void);
+void     engram_connect(el_val_t from_id, el_val_t to_id, el_val_t weight, el_val_t relation);
+el_val_t engram_edge_between(el_val_t from_id, el_val_t to_id);
+el_val_t engram_neighbors(el_val_t node_id);
+el_val_t engram_neighbors_filtered(el_val_t node_id, el_val_t max_depth, el_val_t direction);
+el_val_t engram_neighbors_json(el_val_t node_id, el_val_t max_depth, el_val_t direction);
+el_val_t engram_load(el_val_t path);
+el_val_t engram_save(el_val_t path);
+
 /* ── Private allocator ───────────────────────────────────────────────────── */
 /*
  * el_seed.c carries its own arena for per-request allocation tracking.
@@ -72,10 +148,17 @@ static void seed_request_start(void) {
     _seed_arena_on    = 1;
 }
 
+/* Defined in el_runtime.c. The string-length cache there keys on pointer +
+ * generation; anything that frees or mutates a runtime string must bump the
+ * generation or a reused address could return a stale length. Weak so this
+ * file still links on its own. */
+__attribute__((weak)) void el_str_cache_flush(void);
+
 static void seed_request_end(void) {
     _seed_arena_on = 0;
     for (size_t i = 0; i < _seed_arena.count; i++) free(_seed_arena.ptrs[i]);
     _seed_arena.count = 0;
+    if (el_str_cache_flush) el_str_cache_flush();   /* freed pointers may be reused */
 }
 
 /* el_request_start / el_request_end — formerly defined in el_runtime.c.
@@ -137,6 +220,7 @@ el_val_t __str_set_char(el_val_t s, el_val_t i, el_val_t c) {
     int64_t idx = (int64_t)i;
     if (idx < 0 || idx >= len) return s;
     p[idx] = (char)(unsigned char)(int64_t)c;
+    if (el_str_cache_flush) el_str_cache_flush();  /* in-place write can move the NUL */
     return s;
 }
 
@@ -831,6 +915,219 @@ void __mutex_unlock(el_val_t m) {
     pthread_mutex_unlock(&_el_mutexes[slot]);
 }
 
+/* ── Channels ─────────────────────────────────────────────────────────────── *
+ * Buffered MPMC channel backed by a mutex + condvar + circular buffer.
+ * Ported from the pre-restructure el_runtime.c (b2aac4b) — runtime/channel.el
+ * has always called these five primitives, but they were never carried
+ * forward into el_seed.c when el_runtime.c was consolidated onto the
+ * canonical release copy. Native channels were silently unlinkable on dev
+ * until this port.
+ *
+ * __channel_new(capacity) -> Int (handle)
+ * __channel_send(ch, msg)         — blocks if full (capacity > 0) or never (unbounded)
+ * __channel_recv(ch) -> String    — blocks until a message is available
+ * __channel_try_recv(ch) -> String — non-blocking, returns "" if empty
+ * __channel_close(ch)             — signal no more sends; recv drains remaining
+ *
+ * Bounded channels (cap > 0): circular buffer, sender blocks when full.
+ * Unbounded channels (cap == 0): dynamic array, sender never blocks.
+ */
+#define EL_CHANNEL_MAX 64
+#define EL_CHANNEL_BUF 1024
+
+typedef struct {
+    char**          buf;
+    int             cap;     /* 0 = unbounded (grows dynamically) */
+    int             head, tail, count;
+    int             dyn_cap; /* allocated slots for unbounded mode */
+    int             closed;
+    pthread_mutex_t mu;
+    pthread_cond_t  not_empty;
+    pthread_cond_t  not_full;
+} ElChannel;
+
+static ElChannel       _channels[EL_CHANNEL_MAX];
+static int             _channel_count = 0;
+static pthread_mutex_t _channel_alloc_mu = PTHREAD_MUTEX_INITIALIZER;
+
+el_val_t __channel_new(el_val_t capacity_v) {
+    int cap = (int)(int64_t)capacity_v;
+    if (cap < 0) cap = 0;
+
+    pthread_mutex_lock(&_channel_alloc_mu);
+    if (_channel_count >= EL_CHANNEL_MAX) {
+        pthread_mutex_unlock(&_channel_alloc_mu);
+        fprintf(stderr, "[__channel_new] channel table full\n");
+        return EL_INT(-1);
+    }
+    int slot = _channel_count++;
+    pthread_mutex_unlock(&_channel_alloc_mu);
+
+    ElChannel* ch = &_channels[slot];
+    memset(ch, 0, sizeof(*ch));
+    ch->cap     = cap;
+    ch->closed  = 0;
+    ch->head    = 0;
+    ch->tail    = 0;
+    ch->count   = 0;
+
+    if (cap > 0) {
+        /* Bounded: fixed circular buffer. */
+        ch->buf     = (char**)malloc((size_t)cap * sizeof(char*));
+        ch->dyn_cap = cap;
+    } else {
+        /* Unbounded: start with EL_CHANNEL_BUF slots, grow as needed. */
+        ch->buf     = (char**)malloc(EL_CHANNEL_BUF * sizeof(char*));
+        ch->dyn_cap = EL_CHANNEL_BUF;
+    }
+    if (!ch->buf) {
+        fprintf(stderr, "[__channel_new] out of memory\n");
+        return EL_INT(-1);
+    }
+
+    pthread_mutex_init(&ch->mu, NULL);
+    pthread_cond_init(&ch->not_empty, NULL);
+    pthread_cond_init(&ch->not_full, NULL);
+
+    return EL_INT(slot);
+}
+
+el_val_t __channel_send(el_val_t ch_v, el_val_t msg_v) {
+    int slot = (int)(int64_t)ch_v;
+    if (slot < 0 || slot >= EL_CHANNEL_MAX) return EL_STR("");
+    ElChannel* ch = &_channels[slot];
+
+    const char* msg = EL_CSTR(msg_v);
+    if (!msg) msg = "";
+    char* copy = strdup(msg); /* channel owns the string */
+
+    pthread_mutex_lock(&ch->mu);
+
+    if (ch->closed) {
+        /* Send on closed channel is a no-op (drop the message). */
+        pthread_mutex_unlock(&ch->mu);
+        free(copy);
+        return EL_STR("");
+    }
+
+    if (ch->cap > 0) {
+        /* Bounded: block while full. */
+        while (ch->count >= ch->cap && !ch->closed) {
+            pthread_cond_wait(&ch->not_full, &ch->mu);
+        }
+        if (ch->closed) {
+            pthread_mutex_unlock(&ch->mu);
+            free(copy);
+            return EL_STR("");
+        }
+        ch->buf[ch->tail] = copy;
+        ch->tail = (ch->tail + 1) % ch->cap;
+        ch->count++;
+    } else {
+        /* Unbounded: grow the buffer if needed. */
+        if (ch->count >= ch->dyn_cap) {
+            int new_cap = ch->dyn_cap * 2;
+            char** grown = (char**)realloc(ch->buf, (size_t)new_cap * sizeof(char*));
+            if (!grown) {
+                pthread_mutex_unlock(&ch->mu);
+                free(copy);
+                fprintf(stderr, "[__channel_send] out of memory growing channel\n");
+                return EL_STR("");
+            }
+            /* The circular buffer may have wrapped. Linearise it first.
+             * In unbounded mode head is always 0 (we append at tail, drain
+             * from head), so a simple memmove isn't needed — but if the
+             * buffer did wrap (tail < head after growth), we need to fix up.
+             * Simplest safe path: if tail wrapped, move the head..old_cap
+             * segment to new_cap..new_cap+(old_cap-head). */
+            if (ch->tail < ch->head) {
+                /* Wrapped: [head..old_cap) is the front, [0..tail) is the back. */
+                int front = ch->dyn_cap - ch->head;
+                memmove(grown + ch->dyn_cap, grown + ch->head, (size_t)front * sizeof(char*));
+                ch->head = ch->dyn_cap;
+            }
+            ch->buf     = grown;
+            ch->dyn_cap = new_cap;
+        }
+        ch->buf[ch->tail] = copy;
+        ch->tail = (ch->tail + 1) % ch->dyn_cap;
+        ch->count++;
+    }
+
+    pthread_cond_signal(&ch->not_empty);
+    pthread_mutex_unlock(&ch->mu);
+    return EL_STR("");
+}
+
+el_val_t __channel_recv(el_val_t ch_v) {
+    int slot = (int)(int64_t)ch_v;
+    if (slot < 0 || slot >= EL_CHANNEL_MAX) return EL_STR("");
+    ElChannel* ch = &_channels[slot];
+
+    pthread_mutex_lock(&ch->mu);
+
+    /* Block until there is a message or the channel is closed and drained. */
+    while (ch->count == 0 && !ch->closed) {
+        pthread_cond_wait(&ch->not_empty, &ch->mu);
+    }
+
+    if (ch->count == 0) {
+        /* Closed and empty — signal EOF. */
+        pthread_mutex_unlock(&ch->mu);
+        return EL_STR("");
+    }
+
+    int buf_cap = (ch->cap > 0) ? ch->cap : ch->dyn_cap;
+    char* msg = ch->buf[ch->head];
+    ch->head  = (ch->head + 1) % buf_cap;
+    ch->count--;
+
+    pthread_cond_signal(&ch->not_full);
+    pthread_mutex_unlock(&ch->mu);
+
+    /* Hand the string to the arena so it is freed after the request. */
+    seed_arena_track(msg);
+    return EL_STR(msg);
+}
+
+el_val_t __channel_try_recv(el_val_t ch_v) {
+    int slot = (int)(int64_t)ch_v;
+    if (slot < 0 || slot >= EL_CHANNEL_MAX) return EL_STR("");
+    ElChannel* ch = &_channels[slot];
+
+    pthread_mutex_lock(&ch->mu);
+
+    if (ch->count == 0) {
+        pthread_mutex_unlock(&ch->mu);
+        return EL_STR("");
+    }
+
+    int buf_cap = (ch->cap > 0) ? ch->cap : ch->dyn_cap;
+    char* msg = ch->buf[ch->head];
+    ch->head  = (ch->head + 1) % buf_cap;
+    ch->count--;
+
+    pthread_cond_signal(&ch->not_full);
+    pthread_mutex_unlock(&ch->mu);
+
+    seed_arena_track(msg);
+    return EL_STR(msg);
+}
+
+el_val_t __channel_close(el_val_t ch_v) {
+    int slot = (int)(int64_t)ch_v;
+    if (slot < 0 || slot >= EL_CHANNEL_MAX) return EL_STR("");
+    ElChannel* ch = &_channels[slot];
+
+    pthread_mutex_lock(&ch->mu);
+    ch->closed = 1;
+    /* Wake all blocked recvers and senders so they can observe the close. */
+    pthread_cond_broadcast(&ch->not_empty);
+    pthread_cond_broadcast(&ch->not_full);
+    pthread_mutex_unlock(&ch->mu);
+    return EL_STR("");
+}
+
 /* ── Subprocess ──────────────────────────────────────────────────────────── */
 
 el_val_t __exec(el_val_t cmd) {
@@ -1082,6 +1379,21 @@ el_val_t __engram_scan_nodes_json(el_val_t limit, el_val_t offset) {
     return engram_scan_nodes_json(limit, offset);
 }
 
+el_val_t engram_edges_json(el_val_t limit, el_val_t offset);
+el_val_t __engram_edges_json(el_val_t limit, el_val_t offset) {
+    return engram_edges_json(limit, offset);
+}
+
+el_val_t engram_pool_stats_json(void);
+el_val_t __engram_pool_stats_json(void) { return engram_pool_stats_json(); }
+
+el_val_t el_alloc_count(void);
+el_val_t el_alloc_bytes(void);
+el_val_t el_peak_rss(void);
+el_val_t __el_alloc_count(void) { return el_alloc_count(); }
+el_val_t __el_alloc_bytes(void) { return el_alloc_bytes(); }
+el_val_t __el_peak_rss(void)    { return el_peak_rss(); }
+
 el_val_t __engram_scan_nodes_by_type_json(el_val_t node_type, el_val_t limit, el_val_t offset) {
     return engram_scan_nodes_by_type_json(node_type, limit, offset);
 }
@@ -1094,7 +1406,27 @@ el_val_t __engram_activate_json(el_val_t query, el_val_t depth) {
     return engram_activate_json(query, depth);
 }
 
+/* Forward decls for el_runtime.c symbols this file wraps. el_seed.c does not
+ * include el_runtime.h (documented in lang/AGENTS.md), so each wrapped symbol
+ * needs a prototype here or clang treats it as an implicit declaration (error
+ * under C99+) and the ABI mis-truncates the el_val_t return. */
+el_val_t engram_op_assert_json(el_val_t node_id, el_val_t depth);
+el_val_t engram_node_full_in(el_val_t purview, el_val_t content, el_val_t node_type, el_val_t label,
+                             el_val_t salience, el_val_t importance, el_val_t confidence,
+                             el_val_t tier, el_val_t tags);
+void     engram_connect_in(el_val_t purview, el_val_t from_id, el_val_t to_id,
+                           el_val_t weight, el_val_t relation);
+
 el_val_t __engram_stats_json(void)                         { return engram_stats_json(); }
+el_val_t __engram_op_assert_json(el_val_t node_id, el_val_t depth) { return engram_op_assert_json(node_id, depth); }
+el_val_t __engram_node_full_in(el_val_t purview, el_val_t content, el_val_t node_type, el_val_t label,
+                               el_val_t salience, el_val_t importance, el_val_t confidence,
+                               el_val_t tier, el_val_t tags) {
+    return engram_node_full_in(purview, content, node_type, label, salience, importance, confidence, tier, tags);
+}
+void __engram_connect_in(el_val_t purview, el_val_t from_id, el_val_t to_id, el_val_t weight, el_val_t relation) {
+    engram_connect_in(purview, from_id, to_id, weight, relation);
+}
 el_val_t __engram_list_layers_json(void)                   { return engram_list_layers_json(); }
 
 el_val_t __engram_compile_layered_json(el_val_t intent, el_val_t depth) {

lang/runtime/el_seed.h

@@ -139,6 +139,13 @@ el_val_t  __mutex_new(void);
 void      __mutex_lock(el_val_t m);
 void      __mutex_unlock(el_val_t m);
 
+/* Buffered MPMC channel (runtime/channel.el). capacity=0 means unbounded. */
+el_val_t  __channel_new(el_val_t capacity);
+el_val_t  __channel_send(el_val_t ch, el_val_t msg);     /* blocks if bounded+full */
+el_val_t  __channel_recv(el_val_t ch);                   /* blocks until available */
+el_val_t  __channel_try_recv(el_val_t ch);                /* non-blocking, "" if empty */
+el_val_t  __channel_close(el_val_t ch);
+
 /* ── Subprocess ──────────────────────────────────────────────────────────── */
 
 el_val_t  __exec(el_val_t cmd);            /* popen, capture all stdout, return String */
@@ -233,6 +240,12 @@ el_val_t  __engram_scan_nodes_by_type_json(el_val_t node_type, el_val_t limit, e
 el_val_t  __engram_neighbors_json(el_val_t node_id, el_val_t max_depth, el_val_t direction);
 el_val_t  __engram_activate_json(el_val_t query, el_val_t depth);
 el_val_t  __engram_stats_json(void);
+el_val_t  __engram_op_assert_json(el_val_t node_id, el_val_t depth);
+el_val_t  __engram_node_full_in(el_val_t purview, el_val_t content, el_val_t node_type, el_val_t label,
+                                el_val_t salience, el_val_t importance, el_val_t confidence,
+                                el_val_t tier, el_val_t tags);
+void      __engram_connect_in(el_val_t purview, el_val_t from_id, el_val_t to_id,
+                              el_val_t weight, el_val_t relation);
 el_val_t  __engram_list_layers_json(void);
 el_val_t  __engram_compile_layered_json(el_val_t intent, el_val_t depth);
 

lang/runtime/elbench.el

@@ -0,0 +1,256 @@
+// runtime/elbench.el — growth-curve classifier and complexity gate.
+//
+// Given a geometric sweep of input sizes and the measurements taken at each,
+// classify the growth curve and decide whether it violates a declared bound.
+//
+// ── Why this exists ──────────────────────────────────────────────────────────
+//
+// Constant-factor regressions are annoying. Complexity regressions are outages.
+// An O(n) lookup inside an O(n) loop is invisible at n=100 in a unit test and
+// catastrophic at n=100000 in production. el #132 was exactly that: a strlen()
+// inside a per-character accessor, quadratic, shipped for months.
+//
+// ── THREE signals, not one ───────────────────────────────────────────────────
+//
+// The gate fits time AND allocation-count AND allocation-bytes, and fails if
+// ANY of them exceeds its declared curve. This is not belt-and-braces; each
+// signal is blind to a real defect class the others catch:
+//
+//   * A copy-on-write accumulator rebuilding its buffer allocates ONCE per
+//     iteration — count is exactly linear — while bytes go quadratic.
+//     Count alone passes it.
+//   * el #132's strlen-per-character is pure CPU and allocates NOTHING.
+//     Both allocation signals read FLAT. Only time catches it.
+//
+// The deterministic signals (count, bytes) are preferable where they apply:
+// no statistics, correct on the first run, machine-independent. They are
+// simply not sufficient.
+//
+// ── SCOPE LIMIT — read this before trusting a flat curve ─────────────────────
+//
+// The allocation counters track EL-LEVEL allocation only: strings, ElList and
+// ElMap bodies, their backing arrays, copy-on-write clones, and the realloc
+// growth path. malloc inside engram_*.c and inside libcurl is NOT counted.
+//
+// A flat allocation curve over a workload dominated by engram or HTTP calls is
+// therefore NOT evidence of anything. It means "no El-level allocation growth",
+// not "no allocation growth". Gate El-level complexity with this; do not read
+// third-party memory behaviour into it.
+//
+// ── Classification method ────────────────────────────────────────────────────
+//
+// Sizes must form a geometric sweep (each n double the last). On such a sweep
+// the ratio between consecutive measurements IS the growth exponent, directly:
+//
+//     O(1)  -> 1.0     O(log n) -> ~1.1     O(n) -> 2.0
+//     O(n log n) -> ~2.2     O(n^2) -> 4.0     O(n^3) -> 8.0
+//
+// DEVIATION FROM DESIGN.md 6.2, stated plainly: that section specified Google
+// Benchmark's one-parameter least-squares fit over candidate curves. This uses
+// consecutive ratios instead. The sweep is mandated geometric either way, and
+// on a geometric sweep ratios are directly interpretable and need no floating
+// point. The cost is weaker separation between O(n) and O(n log n), which is
+// reported honestly as an ambiguous band rather than guessed at. Least-squares
+// remains the better answer if that band ever needs to be resolved.
+//
+// All arithmetic is fixed-point, scaled by 1000 ("milli-ratio"), so a ratio of
+// 2.0 is 2000. El values are int64; this avoids float-in-list handling.
+
+// Curve identifiers. Ordered by growth — the ordering IS the comparison used
+// by the gate, so an index comparison decides "worse than declared".
+//   0 = O(1)   1 = O(log n)   2 = O(n)   3 = O(n log n)   4 = O(n^2)   5 = O(n^3)
+
+fn elb_curve_name(c: Int) -> String {
+    if c == 0 { return "O(1)" }
+    if c == 1 { return "O(log n)" }
+    if c == 2 { return "O(n)" }
+    if c == 3 { return "O(n log n)" }
+    if c == 4 { return "O(n^2)" }
+    if c == 5 { return "O(n^3)" }
+    return "O(?)"
+}
+
+fn elb_curve_from_name(s: String) -> Int {
+    if str_eq(s, "O(1)") { return 0 }
+    if str_eq(s, "O(log n)") { return 1 }
+    if str_eq(s, "O(n)") { return 2 }
+    if str_eq(s, "O(n log n)") { return 3 }
+    if str_eq(s, "O(n^2)") { return 4 }
+    if str_eq(s, "O(n^3)") { return 5 }
+    return -1
+}
+
+// elb_classify_ratio — map a milli-ratio-per-doubling onto a curve.
+//
+// Bands are deliberately wide at the top (a quadratic measured at 3.4x is
+// still a quadratic) and deliberately overlap-averse at the bottom, where a
+// misclassification between O(1) and O(log n) matters least.
+fn elb_classify_ratio(milli: Int) -> Int {
+    if milli < 1300 { return 0 }
+    if milli < 1700 { return 1 }
+    if milli < 2400 { return 2 }
+    if milli < 3200 { return 3 }
+    if milli < 6000 { return 4 }
+    return 5
+}
+
+// elb_ratio — milli-ratio between two consecutive measurements.
+// Returns -1 when the earlier measurement is zero (ratio undefined).
+fn elb_ratio(prev: Int, cur: Int) -> Int {
+    if prev <= 0 { return -1 }
+    return (cur * 1000) / prev
+}
+
+// ── The measurement floor ────────────────────────────────────────────────────
+//
+// A benchmark whose largest measurement is at or near zero has not been
+// measured. Reporting it as O(1) would be a confident answer with nothing
+// behind it — the same failure as a test that never ran reporting pass, and
+// exactly what happened when clang closed a nested loop to a multiply and the
+// harness read 0 microseconds at every n.
+//
+// So: REFUSE. Never classify below the floor.
+fn elb_below_floor(vals: [Int], floor: Int) -> Bool {
+    let n: Int = native_list_len(vals)
+    let i: Int = 0
+    let mx: Int = 0
+    while i < n {
+        let v: Int = native_list_get(vals, i)
+        if v > mx { let mx = v }
+        let i = i + 1
+    }
+    if mx < floor { return true }
+    return false
+}
+
+// elb_implausibly_flat — a measurement that does not move across a sweep whose
+// input grew by 8x or more is not a flat curve, it is a broken measurement.
+// Genuine O(1) work still shows noise; a hard-flat series means the work was
+// optimised away, the timer has insufficient resolution, or the benchmark body
+// never executed.
+fn elb_implausibly_flat(vals: [Int]) -> Bool {
+    let n: Int = native_list_len(vals)
+    if n < 3 { return false }
+    let first: Int = native_list_get(vals, 0)
+    let last: Int = native_list_get(vals, n - 1)
+    if first == 0 {
+        if last == 0 { return true }
+        return false
+    }
+    let r: Int = (last * 1000) / first
+    if r < 1100 { return true }
+    return false
+}
+
+// elb_spread_ok — do the consecutive ratios agree with each other?
+//
+// This is the ratio-method analogue of a normalised-RMS threshold. If the
+// doublings disagree wildly the data is noise, a cache cliff, or a phase
+// change, and the honest report is INDETERMINATE rather than a classification.
+// Applies to the ASYMPTOTIC TAIL only — the last three ratios.
+//
+// The small-n end of any sweep is dominated by fixed overhead, cold caches and
+// branch predictors that have not warmed. Measured on a genuinely linear
+// character scan, the ratios ran 3.37, 2.92, 1.76, 1.65: the head looks
+// quadratic, the tail is the truth. Checking spread across the whole sweep
+// therefore rejects correct data. A complexity bound is an asymptotic claim, so
+// it is judged on the asymptotic region — the same reason a benchmark harness
+// discards warmup rather than averaging it in.
+fn elb_spread_ok(ratios: [Int]) -> Bool {
+    let total: Int = native_list_len(ratios)
+    if total < 2 { return true }
+    let start: Int = total - 3
+    if start < 0 { let start = 0 }
+    let n: Int = total
+    let lo: Int = 999999
+    let hi: Int = 0
+    let i: Int = start
+    while i < n {
+        let r: Int = native_list_get(ratios, i)
+        if r >= 0 {
+            if r < lo { let lo = r }
+            if r > hi { let hi = r }
+        }
+        let i = i + 1
+    }
+    if lo <= 0 { return false }
+    // Reject when the widest ratio is more than 2.2x the narrowest. That is
+    // enough slack for real timing noise and tight enough to separate a clean
+    // 2.0 series from a clean 4.0 series.
+    if (hi * 1000) / lo > 2200 { return false }
+    return true
+}
+
+// elb_ratios — consecutive milli-ratios across the sweep.
+fn elb_ratios(vals: [Int]) -> [Int] {
+    let out: [Int] = native_list_empty()
+    let n: Int = native_list_len(vals)
+    let i: Int = 1
+    while i < n {
+        let out = native_list_append(out,
+            elb_ratio(native_list_get(vals, i - 1), native_list_get(vals, i)))
+        let i = i + 1
+    }
+    return out
+}
+
+// elb_mean_tail_ratio — mean of the LAST TWO ratios.
+//
+// The tail is used deliberately: asymptotic behaviour is what a complexity
+// bound claims, and the small-n end of any sweep is dominated by fixed
+// overhead. This is the same reason a benchmark harness discards warmup.
+fn elb_mean_tail_ratio(ratios: [Int]) -> Int {
+    let n: Int = native_list_len(ratios)
+    if n == 0 { return -1 }
+    if n == 1 { return native_list_get(ratios, 0) }
+    let a: Int = native_list_get(ratios, n - 1)
+    let b: Int = native_list_get(ratios, n - 2)
+    if a < 0 { return b }
+    if b < 0 { return a }
+    return (a + b) / 2
+}
+
+// ── Verdicts ─────────────────────────────────────────────────────────────────
+//
+//   0 PASS           measured curve is at or below the declared bound
+//   1 FAIL           measured curve is strictly worse than declared
+//   2 INDETERMINATE  ratios disagree; data is noise or a phase change
+//   3 REFUSED        below the measurement floor, or implausibly flat
+//   4 BETTER         measured strictly better than declared (warn, not fail)
+
+fn elb_verdict_name(v: Int) -> String {
+    if v == 0 { return "PASS" }
+    if v == 1 { return "FAIL" }
+    if v == 2 { return "INDETERMINATE" }
+    if v == 3 { return "REFUSED" }
+    if v == 4 { return "BETTER" }
+    return "?"
+}
+
+// elb_gate — classify one signal against its declared bound.
+//
+//   vals     measurements, one per sweep point, in sweep order
+//   expect   declared curve index (see elb_curve_name)
+//   floor    minimum largest-measurement below which we refuse to classify
+fn elb_gate(vals: [Int], expect: Int, floor: Int) -> Int {
+    if elb_below_floor(vals, floor) { return 3 }
+    if elb_implausibly_flat(vals) { return 3 }
+    let ratios: [Int] = elb_ratios(vals)
+    if !elb_spread_ok(ratios) { return 2 }
+    let m: Int = elb_mean_tail_ratio(ratios)
+    if m < 0 { return 2 }
+    let got: Int = elb_classify_ratio(m)
+    if got > expect { return 1 }
+    if got < expect { return 4 }
+    return 0
+}
+
+// elb_measured_curve — the classified curve for a signal, or -1 if unclassifiable.
+fn elb_measured_curve(vals: [Int], floor: Int) -> Int {
+    if elb_below_floor(vals, floor) { return -1 }
+    if elb_implausibly_flat(vals) { return -1 }
+    let ratios: [Int] = elb_ratios(vals)
+    let m: Int = elb_mean_tail_ratio(ratios)
+    if m < 0 { return -1 }
+    return elb_classify_ratio(m)
+}

lang/runtime/eltest.el

@@ -0,0 +1,194 @@
+// runtime/eltest.el — El test framework runner (Phase 1).
+//
+// This is the RUNNER. It is written in El and consumes a registry that the
+// compiler generates into the same translation unit when invoked as
+// `elc --test`. Nothing here discovers tests; discovery already happened at
+// compile time, which is what makes `--list` and filtering possible later.
+//
+// ── Architecture ─────────────────────────────────────────────────────────────
+//
+// The compiler lowers each `test "name" { ... }` block into a static C
+// function and emits a static table of (name, fn) pairs plus a small set of
+// index-based accessors. El has no function pointers, so the runner never
+// sees one — it works entirely in indices:
+//
+//   __el_reg_count()      -> Int     number of registered tests
+//   __el_reg_name(i)      -> String  test name at index i
+//   __el_reg_invoke(i)    -> Int     run test i, return its failure count
+//   __el_reg_last_ns()    -> Int     wall-clock ns of the last invoke
+//   __el_reg_msg()        -> String  first failure message of the last invoke
+//   __el_reg_asserts()    -> Int     assertions executed in the last invoke
+//   __el_opt_json()       -> Int     1 if --json was passed
+//
+// Timing is taken in the generated C, immediately around the call, so no El
+// call overhead lands inside the measurement.
+//
+// ── Output ───────────────────────────────────────────────────────────────────
+//
+// Structured events are the source of truth. The human renderer is written
+// FROM the same fields the NDJSON renderer emits — never the reverse. Parsing
+// human output back into structure is the one clear architectural mistake in
+// Go's test tooling and we do not repeat it.
+//
+// Every result carries a duration. Always. A framework that cannot report how
+// long its tests took cannot surface a performance regression, and a
+// regression nobody can see is one nobody fixes.
+
+// ── Small helpers (no imports — this file must stay self-contained) ──────────
+
+// _elt_json_escape — minimal JSON string escaping for the NDJSON renderer.
+fn _elt_json_escape(s: String) -> String {
+    let out: String = ""
+    let n: Int = str_len(s)
+    let i: Int = 0
+    while i < n {
+        let ch: String = str_slice(s, i, i + 1)
+        if str_eq(ch, "\"") {
+            let out = out + "\\\""
+        } else {
+            if str_eq(ch, "\\") {
+                let out = out + "\\\\"
+            } else {
+                if str_eq(ch, "\n") {
+                    let out = out + "\\n"
+                } else {
+                    if str_eq(ch, "\t") {
+                        let out = out + "\\t"
+                    } else {
+                        if str_eq(ch, "\r") {
+                            let out = out + "\\r"
+                        } else {
+                            let out = out + ch
+                        }
+                    }
+                }
+            }
+        }
+        let i = i + 1
+    }
+    return out
+}
+
+// _elt_pad3 — left-pad an integer to three digits (for the ms.fraction form).
+fn _elt_pad3(v: Int) -> String {
+    if v < 10 { return "00" + int_to_str(v) }
+    if v < 100 { return "0" + int_to_str(v) }
+    return int_to_str(v)
+}
+
+// _elt_ms — render a nanosecond duration as "M.mmm" milliseconds.
+//
+// Deliberately avoids the modulo operator: the remainder is derived by
+// subtraction so this stays portable across El backends.
+fn _elt_ms(ns: Int) -> String {
+    let total_us: Int = ns / 1000
+    let ms_whole: Int = total_us / 1000
+    let us_rem: Int = total_us - (ms_whole * 1000)
+    return int_to_str(ms_whole) + "." + _elt_pad3(us_rem)
+}
+
+// _elt_secs — render a nanosecond duration as fractional seconds, for the
+// NDJSON `elapsed` field. JUnit XML and test2json both use seconds-as-decimal.
+fn _elt_secs(ns: Int) -> String {
+    let total_ms: Int = ns / 1000000
+    let s_whole: Int = total_ms / 1000
+    let ms_rem: Int = total_ms - (s_whole * 1000)
+    return int_to_str(s_whole) + "." + _elt_pad3(ms_rem)
+}
+
+// ── Event emission ───────────────────────────────────────────────────────────
+//
+// One function per event shape. Both renderers read the same fields; the
+// human renderer is a projection of the event, not a separate code path.
+
+fn _elt_emit_run(json_mode: Bool, name: String) {
+    if json_mode {
+        println("{\"action\":\"run\",\"test\":\"" + _elt_json_escape(name) + "\"}")
+    }
+}
+
+fn _elt_emit_result(json_mode: Bool, name: String, fails: Int, ns: Int, asserts: Int, msg: String) {
+    if json_mode {
+        let action: String = "pass"
+        if fails > 0 { let action = "fail" }
+        let line: String = "{\"action\":\"" + action + "\""
+        let line = line + ",\"test\":\"" + _elt_json_escape(name) + "\""
+        let line = line + ",\"elapsed\":" + _elt_secs(ns)
+        let line = line + ",\"assertions\":" + int_to_str(asserts)
+        if fails > 0 {
+            let line = line + ",\"failures\":" + int_to_str(fails)
+            let line = line + ",\"message\":\"" + _elt_json_escape(msg) + "\""
+        }
+        let line = line + "}"
+        println(line)
+        return
+    }
+    // Human renderer — duration is never optional.
+    if fails > 0 {
+        println("FAIL  " + name + "  (" + _elt_ms(ns) + "ms)")
+        println("      " + msg)
+        return
+    }
+    println("ok    " + name + "  (" + _elt_ms(ns) + "ms)")
+    return
+}
+
+fn _elt_emit_summary(json_mode: Bool, total: Int, failed: Int, ns: Int, asserts: Int) {
+    let passed: Int = total - failed
+    if json_mode {
+        let line: String = "{\"action\":\"summary\""
+        let line = line + ",\"tests\":" + int_to_str(total)
+        let line = line + ",\"passed\":" + int_to_str(passed)
+        let line = line + ",\"failed\":" + int_to_str(failed)
+        let line = line + ",\"assertions\":" + int_to_str(asserts)
+        let line = line + ",\"elapsed\":" + _elt_secs(ns)
+        let line = line + "}"
+        println(line)
+        return
+    }
+    println("")
+    println(int_to_str(total) + " tests, " + int_to_str(passed) + " passed, "
+        + int_to_str(failed) + " failed, " + int_to_str(asserts) + " assertions in "
+        + _elt_ms(ns) + "ms")
+    return
+}
+
+// ── The runner ───────────────────────────────────────────────────────────────
+
+// el_test_main — drive the compile-time registry.
+//
+// Called from the generated main(). Returns the number of FAILING TESTS, which
+// becomes the process exit code. Note that this counts tests, not assertions:
+// a test is the unit of result. The old harness counted assertions globally and
+// therefore could not say which test failed, how long any of them took, or
+// whether a test had run at all.
+fn el_test_main() -> Int {
+    let json_mode: Bool = false
+    if __el_opt_json() == 1 { let json_mode = true }
+
+    let n: Int = __el_reg_count()
+    let i: Int = 0
+    let failed: Int = 0
+    let total_ns: Int = 0
+    let total_asserts: Int = 0
+
+    while i < n {
+        let name: String = __el_reg_name(i)
+        _elt_emit_run(json_mode, name)
+
+        let fails: Int = __el_reg_invoke(i)
+        let ns: Int = __el_reg_last_ns()
+        let asserts: Int = __el_reg_asserts()
+        let msg: String = __el_reg_msg()
+
+        let total_ns = total_ns + ns
+        let total_asserts = total_asserts + asserts
+        if fails > 0 { let failed = failed + 1 }
+
+        _elt_emit_result(json_mode, name, fails, ns, asserts, msg)
+        let i = i + 1
+    }
+
+    _elt_emit_summary(json_mode, n, failed, total_ns, total_asserts)
+    return failed
+}

lang/runtime/engram_geometry.c

@@ -222,7 +222,7 @@ static double eff_w(double weight, double hebb){
 }
 
 GeoDescriptor* engram_geometry_descriptor(
-    EngramPagedStore* store, VIndex* vindex,
+    EngramPagedStore* store, const VIndex* vindex,
     char** vids, int n_vids,
     const char* const* seed_ids, size_t n_seeds,
     const GeoParams* params,
@@ -1401,7 +1401,7 @@ static double geo_weighted_degree(EngramPagedStore* st, const char* id, double e
     return deg;
 }
 
-int engram_geo_reify_store(EngramPagedStore* store, VIndex* vindex,
+int engram_geo_reify_store(EngramPagedStore* store, const VIndex* vindex,
                            char** vids, int n_vids,
                            const GeoReifyParams* params){
     if(!store) return -1;

lang/runtime/engram_geometry.h

@@ -150,7 +150,7 @@ void engram_geo_mean_free(GeoMeanCache* c);
  * Returns a malloc'd descriptor (free with engram_geo_free), or NULL on error
  * (no seeds resolvable, OOM). */
 GeoDescriptor* engram_geometry_descriptor(
-    EngramPagedStore* store, VIndex* vindex,
+    EngramPagedStore* store, const VIndex* vindex,
     char** vids, int n_vids,
     const char* const* seed_ids, size_t n_seeds,
     const GeoParams* params,
@@ -375,7 +375,7 @@ void engram_geo_reify_default_params(GeoReifyParams* p);
  * neighborhood (+ member edges), superseding any prior same-hub record with
  * provenance. Read-then-write over `store`. Returns #neighborhoods persisted, or <0.
  * Skips existing Neighborhood/GeoMeanFrame nodes when detecting (idempotent re-reify). */
-int engram_geo_reify_store(EngramPagedStore* store, VIndex* vindex,
+int engram_geo_reify_store(EngramPagedStore* store, const VIndex* vindex,
                            char** vids, int n_vids,
                            const GeoReifyParams* params);
 

lang/runtime/engram_store.c

@@ -44,6 +44,11 @@
 #include <string.h>
 #include <stdint.h>
 #include <unistd.h>
+#if defined(__APPLE__) || defined(__MACH__)
+#include <sys/sysctl.h>
+#include <mach/mach.h>
+#include <mach/mach_host.h>
+#endif
 #include <fcntl.h>
 #include <errno.h>
 #include <time.h>
@@ -236,8 +241,16 @@ struct PgCache {
     unsigned prefetch;           /* read-ahead window (pages); 0 = off */
     LayerPin* lp; size_t lp_n, lp_cap;   /* hot-layer pin bookkeeping */
     size_t   dirty_count;        /* # dirty frames, maintained incrementally (M5) */
-    /* stats (introspection only — never affect semantics) */
+    /* Interoception. These were "introspection only — never affect semantics",
+     * and that was the bug: the pool could not feel itself thrash, so it could
+     * not correct, and neither could anyone watching from outside. The sensed
+     * state IS the corrective mechanism (see pc_adapt_budget) — the same way the
+     * engram's own boundary-beat/chronoception let it feel its own activity. */
     uint64_t hits, misses, evictions, prefetch_reads;
+    /* sliding-window marks so pressure reflects NOW, not lifetime totals */
+    uint64_t adapt_last_acc, adapt_last_evic, adapt_last_hits;
+    uint64_t adapt_grows;        /* budget corrections upward   */
+    uint64_t adapt_shrinks;      /* budget corrections downward (memory pressure) */
 };
 
 /* ── little-endian scalar codecs ──────────────────────────────────────────── */
@@ -334,6 +347,51 @@ static uint64_t dh_node_hash(const StoreNode* n){
     return h;
 }
 
+/* dh_edge_hash — the edge counterpart of dh_node_hash.
+ *
+ * WHY THIS EXISTS (2026-08-15): the write barrier was node-only. Checkpointing
+ * pushes the WHOLE resident graph through store_put_node/store_put_edge (see
+ * engram_store_checkpoint), and nodes were cheaply skipped when unchanged —
+ * a hash compare, no page I/O. Edges had no such check, so every edge was
+ * rewritten on every checkpoint, and each rewrite runs the idempotency probe
+ * max_page_lsn_for_id → btree lookup → page_read per stored copy.
+ *
+ * Edges outnumber nodes roughly 3:1 here (37,663 vs 13,430), so this turned
+ * routine checkpointing into a FULL-STORE WALK in id order — random page access
+ * across the entire 2 GiB store, repeated, mostly to rediscover that nothing
+ * had changed. That walk is the failure mode: with a page cache smaller than
+ * the store it degenerates into thrashing and the engram never makes progress.
+ * Sizing the cache around that walk treats the symptom; the walk itself should
+ * not happen.
+ *
+ * The discriminator byte keeps the edge keyspace from ever colliding with a
+ * node of the same id in the shared dh map: distinct kinds cannot produce the
+ * same hash, so a stale skip is not reachable by collision. */
+static uint64_t dh_edge_hash(const StoreEdge* e){
+    uint64_t h = 1469598103934665603ULL;
+    const uint8_t kind = 0xE0;           /* edge discriminator */
+    dh_fold_bytes(&h, &kind, 1);
+    dh_fold_str(&h, e->id);
+    dh_fold_str(&h, e->from_id);
+    dh_fold_str(&h, e->to_id);
+    dh_fold_str(&h, e->relation);
+    dh_fold_str(&h, e->metadata);
+    uint8_t t8[8];
+    put_f64(t8, e->weight);     dh_fold_bytes(&h, t8, 8);
+    put_f64(t8, e->hebb);       dh_fold_bytes(&h, t8, 8);
+    put_f64(t8, e->confidence); dh_fold_bytes(&h, t8, 8);
+    uint8_t t4[4];
+    put_u32(t4, (uint32_t)e->inhibitory); dh_fold_bytes(&h, t4, 4);
+    put_u32(t4, e->layer_id);             dh_fold_bytes(&h, t4, 4);
+    /* created_at/updated_at/last_fired are deliberately EXCLUDED: last_fired is
+     * touched by activation without changing what the edge IS, and including it
+     * would defeat the barrier on exactly the hot edges it most needs to skip.
+     * The fields that define the edge's durable content are all folded above. */
+    if (e->unknown && e->unknown_len) dh_fold_bytes(&h, e->unknown, e->unknown_len);
+    if (h == 0) h = 1;   /* reserve 0 as "absent" in the map */
+    return h;
+}
+
 /* Open-addressing id(string)→durable-hash map. Keyed for O(1) bucketing on the
  * id's FNV hash, compared by strcmp for correctness (full-id discipline, matching
  * store_scan_*'s StrSet). Values are the 64-bit durable hash. */
@@ -1531,6 +1589,11 @@ int store_scan_edges(EngramPagedStore* s, StoreEdgeScanCb cb, void* ctx){
             if (cand.id && *cand.id && strset_add(&seen, cand.id)){
                 StoreEdge canon;
                 if (store_get_edge(s, cand.id, &canon) == 1){
+                    /* seed the write-barrier map from on-disk truth so the FIRST
+                     * post-boot checkpoint full-walk already skips unchanged edges
+                     * (mirrors store_scan_nodes; without it the barrier is empty at
+                     * boot and the first checkpoint re-probes every edge) */
+                    if (s->barrier_on) dh_set(s->dh, canon.id, dh_edge_hash(&canon));
                     cb(&canon, ctx); count++;               /* canonical latest-live */
                     store_edge_free(&canon);
                 }
@@ -1594,19 +1657,76 @@ int store_scan_edges(EngramPagedStore* s, StoreEdgeScanCb cb, void* ctx){
  *  matches disk, so a re-fault reproduces identical bytes.
  * ════════════════════════════════════════════════════════════════════════════ */
 
-/* default frame budget: large enough that today's whole store stays resident
- * (== Phase 1). Override with env ENGRAM_POOL_FRAMES (0 = unlimited). */
-#ifndef ENGRAM_POOL_FRAMES_DEFAULT
-#define ENGRAM_POOL_FRAMES_DEFAULT (1u<<20)   /* ~1M frames × 16KiB = 16 GiB */
+/* ── Frame budget ────────────────────────────────────────────────────────────
+ *
+ * A FIXED frame count cannot be correct. It has no relationship to either
+ * quantity that decides whether a cache works: the size of the working set, or
+ * the memory actually available on the host. It is the same number on a 16 GB
+ * laptop and a 256 GB server, and it stays put while the store grows.
+ *
+ * That is not hypothetical. On 2026-08-15 the deployment pinned
+ * ENGRAM_POOL_FRAMES=65536 (1 GiB) while neuron.egm grew to 2.1 GiB. The
+ * working set was twice the budget, so boot-time WAL replay — which walks
+ * pages in an order uncorrelated with reuse — evicted each page shortly before
+ * it was needed again. The engram spun at 100% CPU inside pc_evict_to_budget
+ * and never bound its port. Not slow: making no progress. Denning's thrashing,
+ * exactly, and no eviction policy can fix it — when the working set does not
+ * fit, only more frames or admission control help.
+ *
+ * So the budget is DERIVED, from the host's physical memory, and it scales
+ * with the machine instead of pretending memory is a constant.
+ *
+ *   ENGRAM_POOL_FRAMES   explicit frame count; 0 = unlimited. Overrides all.
+ *                        Prefer leaving it unset — a hand-set number is how
+ *                        this failure happened.
+ *   ENGRAM_POOL_MEM_PCT  percent of physical RAM to budget (default 60).
+ *
+ * Fallback when RAM cannot be read is 16 GiB worth of frames — the old
+ * default, retained only as a floor for that case.
+ * ──────────────────────────────────────────────────────────────────────────── */
+#ifndef ENGRAM_POOL_FRAMES_FALLBACK
+#define ENGRAM_POOL_FRAMES_FALLBACK (1u<<20)   /* ~1M frames × 16KiB = 16 GiB */
 #endif
 
+static uint64_t pc_available_ram(void);   /* fwd — defined with the controller */
+
+/* Physical RAM in bytes, 0 when it cannot be determined. */
+static uint64_t pc_physical_ram(void){
+#if defined(__APPLE__) || defined(__MACH__)
+    uint64_t v = 0; size_t len = sizeof v;
+    int mib[2] = { CTL_HW, HW_MEMSIZE };
+    if (sysctl(mib, 2, &v, &len, NULL, 0) == 0) return v;
+    return 0;
+#else
+    long pages = sysconf(_SC_PHYS_PAGES);
+    long psz   = sysconf(_SC_PAGESIZE);
+    if (pages > 0 && psz > 0) return (uint64_t)pages * (uint64_t)psz;
+    return 0;
+#endif
+}
+
+static size_t pc_default_cap(void){
+    unsigned pct = 60;
+    const char* p = getenv("ENGRAM_POOL_MEM_PCT");
+    if (p && *p){ unsigned long v = strtoul(p, NULL, 10); if (v > 0 && v <= 95) pct = (unsigned)v; }
+    uint64_t ram = pc_physical_ram();
+    if (!ram) return ENGRAM_POOL_FRAMES_FALLBACK;
+    uint64_t budget_bytes = (ram / 100u) * pct;
+    /* Never start above what the machine can actually spare right now. */
+    uint64_t avail = pc_available_ram();
+    if (avail > (1ull<<30) && budget_bytes > avail - (1ull<<30)) budget_bytes = avail - (1ull<<30);
+    uint64_t frames = budget_bytes / (uint64_t)STORE_PAGE_SIZE;
+    if (frames < 4096) frames = 4096;            /* never absurdly small */
+    return (size_t)frames;
+}
+
 static PgCache* pc_new(void){
     PgCache* c = (PgCache*)calloc(1, sizeof *c);
     if (!c) return NULL;
     c->nbuckets = 1024;
     c->buckets = (PgEnt**)calloc(c->nbuckets, sizeof(PgEnt*));
     if (!c->buckets){ free(c); return NULL; }
-    c->cap = ENGRAM_POOL_FRAMES_DEFAULT;
+    c->cap = pc_default_cap();
     c->prefetch = 8;
     const char* pf = getenv("ENGRAM_POOL_FRAMES");
     if (pf && *pf){ char* end=NULL; unsigned long long v = strtoull(pf,&end,10); c->cap = (size_t)v; }
@@ -1676,6 +1796,241 @@ static void pc_remove(PgCache* c, PgEnt* e){
 /* Reclaim clean unpinned frames from the LRU end until under budget, or until no
  * evictable frame remains (a dirty/pinned-heavy pool may transiently exceed cap —
  * that is the no-steal guarantee, not a bug: the next checkpoint frees them). */
+/* ── Adaptive budget: close the loop ─────────────────────────────────────────
+ *
+ * THE LESSON THIS ENCODES (2026-08-15). The engram spent hours down while four
+ * separate theories were tried — bad binary, corrupt snapshot, WAL replay,
+ * feature flags — because nothing in the system said what was happening. It
+ * looked identical to "busy loading": 100% CPU, flat RSS, no output. Meanwhile
+ * hits/misses/evictions were ALREADY being counted, right here, and surfaced
+ * nowhere. One eviction-rate number would have ended it in seconds.
+ *
+ * So the counters are not decoration. They are the control signal.
+ *
+ * A budget chosen once — a literal like 65536, or 60% of RAM read at startup —
+ * is a guess about the future. It cannot know the store grew, the working set
+ * shifted, or another process took the memory. The cache already MEASURES the
+ * only thing that matters (am I evicting pages I am about to want again), so it
+ * should act on that measurement instead of on a number someone typed.
+ *
+ * The controller: over a sliding window, if evictions are running at a rate
+ * comparable to accesses AND there is genuine reuse (hits are material), the
+ * working set exceeds the budget — grow it. Growth is geometric, bounded by a
+ * live re-read of physical memory rather than a value cached at boot, so it
+ * tracks the machine instead of a snapshot of it. It never shrinks on its own:
+ * cap is a ceiling, not an allocation, and frames are only ever held because a
+ * real access put them there.
+ *
+ * Two things this deliberately does NOT do: it does not attempt a cleverer
+ * eviction policy (when the working set does not fit, no policy helps — that is
+ * Denning, and it is why "tune the LRU" was never the fix), and it does not stay
+ * silent (pool_report exposes the same numbers outward, so a human or a metric
+ * pipeline sees the pressure the controller is reacting to). */
+
+/* El's native telemetry, already in the runtime and already exporting to OTLP.
+ * Declared weak so engram_store.c still links standalone; when the runtime is
+ * present (every real build) the pool's interoception flows into the SAME
+ * pipeline as every other metric.
+ *
+ * ONE emission carrying the whole sensed state — not a function per stat, and
+ * not a bespoke per-subsystem endpoint. Both of those are the degenerate case:
+ * they make observability something you hand-write per noun instead of a
+ * uniform mechanism every component already has. el_val_t is int64_t; strings
+ * ride as pointers cast through it (see el_runtime.h's value model). */
+__attribute__((weak)) int64_t emit_log(int64_t level, int64_t msg, int64_t fields_json);
+
+static void pc_report(const PgCache* c, const char* cause){
+    if (!emit_log) return;                    /* runtime not linked: no-op */
+    uint64_t acc = c->hits + c->misses;
+    char f[512];
+    snprintf(f, sizeof f,
+        "{\"component\":\"engram.pool\",\"cause\":\"%s\",\"hits\":%llu,\"misses\":%llu,"
+        "\"evictions\":%llu,\"prefetch_reads\":%llu,\"cap_frames\":%zu,\"resident\":%zu,"
+        "\"dirty\":%zu,\"grows\":%llu,\"hit_rate\":%.4f,\"evict_ratio\":%.4f,"
+        "\"cap_gib\":%.3f,\"resident_gib\":%.3f}",
+        cause,
+        (unsigned long long)c->hits, (unsigned long long)c->misses,
+        (unsigned long long)c->evictions, (unsigned long long)c->prefetch_reads,
+        c->cap, c->count, c->dirty_count, (unsigned long long)c->adapt_grows,
+        acc ? (double)c->hits / (double)acc : 0.0,
+        acc ? (double)c->evictions / (double)acc : 0.0,
+        (double)c->cap   * (double)STORE_PAGE_SIZE / (1024.0*1024.0*1024.0),
+        (double)c->count * (double)STORE_PAGE_SIZE / (1024.0*1024.0*1024.0));
+    emit_log((int64_t)(uintptr_t)"warn", (int64_t)(uintptr_t)"engram.pool pressure",
+             (int64_t)(uintptr_t)f);
+}
+
+static uint64_t pc_ram_bytes_live(void){ return pc_physical_ram(); }
+
+/* AVAILABLE memory right now — free + reclaimable, not total.
+ *
+ * Sizing a cache against TOTAL ram is what turns a cache into a memory leak:
+ * total does not shrink when other processes need memory, so a pool that only
+ * grows never notices it is starving the machine it runs on. Availability does.
+ * Returns 0 when undeterminable — callers then refuse to grow, the safe way. */
+static uint64_t pc_available_ram(void){
+#if defined(__APPLE__) || defined(__MACH__)
+    /* SWAP AND COMPRESSOR FIRST. free+inactive+purgeable is a LIE under memory
+     * pressure: a machine deep in swap still reports gigabytes "available",
+     * because inactive pages are only reclaimable by evicting them to swap.
+     * Observed 2026-08-15: this returned 9.43 GiB available while vm.swapusage
+     * showed 51.58 of 53.25 GiB used (97% full) and the compressor occupied
+     * 23.7 GiB — the host was thrashing to disk and the pool would have been
+     * cleared to grow into it. Growing a cache in that state is how a guard
+     * becomes the crash.
+     *
+     * So: if swap is nearly spent, report ZERO available. Callers refuse to
+     * grow on 0 and pc_relieve_pressure hands frames back. Only when the
+     * machine is genuinely not swapping do free+inactive+purgeable mean
+     * anything, and even then the compressor's footprint is subtracted because
+     * that RAM is already spoken for. */
+    /* RATE, NOT LEVEL. Swap *level* is a terrible signal: macOS grows swap files
+     * on demand and reclaims them lazily, so "47 of 48 GiB used" can mean the
+     * machine is dying OR that it recovered ten minutes ago and the file has not
+     * been trimmed yet. Measured both states on one host within minutes:
+     *   47.65/48.00 GiB used, 2047 swapouts/s  -> genuinely thrashing
+     *   26.67/28.00 GiB used,    0 swapouts/s  -> perfectly healthy, 15.6 GiB free
+     * A level check calls the second one an emergency and starves the pool for
+     * no reason. What distinguishes them is whether pages are moving NOW.
+     *
+     * So sample the swapout counter across calls and judge the delta. First call
+     * establishes the baseline and reports no pressure — one sample cannot have
+     * a rate, and guessing from a single reading is the whole mistake. */
+    {
+        static uint64_t prev_swapouts = 0;
+        static time_t   prev_t = 0;
+        static int      primed = 0;
+        mach_port_t h0 = mach_host_self();
+        vm_statistics64_data_t v0; mach_msg_type_number_t c0 = HOST_VM_INFO64_COUNT;
+        if (host_statistics64(h0, HOST_VM_INFO64, (host_info64_t)&v0, &c0) == KERN_SUCCESS){
+            uint64_t now_out = (uint64_t)v0.swapouts;
+            time_t   now_t   = time(NULL);
+            if (!primed){ prev_swapouts = now_out; prev_t = now_t; primed = 1; }
+            else if (now_t > prev_t){
+                double per_s = (double)(now_out - prev_swapouts) / (double)(now_t - prev_t);
+                prev_swapouts = now_out; prev_t = now_t;
+                /* Sustained outward paging with nothing coming back is the
+                 * signature of a host being pushed into swap. ~200 pages/s is
+                 * ~3 MiB/s — well above idle noise, well below the 2000+/s seen
+                 * while actually thrashing. */
+                if (per_s > 200.0) return 0;
+            }
+        }
+    }
+    mach_port_t host = mach_host_self();
+    vm_size_t page = 0;
+    if (host_page_size(host, &page) != KERN_SUCCESS) return 0;
+    vm_statistics64_data_t vm; mach_msg_type_number_t cnt = HOST_VM_INFO64_COUNT;
+    if (host_statistics64(host, HOST_VM_INFO64, (host_info64_t)&vm, &cnt) != KERN_SUCCESS) return 0;
+    uint64_t avail = (uint64_t)vm.free_count + (uint64_t)vm.inactive_count
+                   + (uint64_t)vm.purgeable_count;
+    /* the compressor is holding real RAM that nobody can hand us */
+    uint64_t compressed = (uint64_t)vm.compressor_page_count;
+    if (compressed >= avail) return 0;
+    avail -= compressed;
+    return avail * (uint64_t)page;
+#else
+    FILE* f = fopen("/proc/meminfo", "r");
+    if (!f) return 0;
+    char line[256]; unsigned long long kb = 0;
+    while (fgets(line, sizeof line, f))
+        if (sscanf(line, "MemAvailable: %llu kB", &kb) == 1) break;
+    fclose(f);
+    return (uint64_t)kb * 1024ull;
+#endif
+}
+
+/* Shrink the budget when the machine is short on memory.
+ *
+ * A pool that can only grow is a leak with extra steps. This is the other half
+ * of the control loop: if free memory drops below a floor, hand frames back.
+ * The resident set follows on the next eviction pass, so the memory is actually
+ * returned rather than merely re-labelled. */
+#ifndef ENGRAM_POOL_FREE_FLOOR_BYTES
+#define ENGRAM_POOL_FREE_FLOOR_BYTES (2ull*1024ull*1024ull*1024ull)   /* 2 GiB */
+#endif
+static int pc_relieve_pressure(PgCache* c){
+    uint64_t avail = pc_available_ram();
+    if (!avail) return 0;
+    uint64_t floor_b = ENGRAM_POOL_FREE_FLOOR_BYTES;
+    const char* fe = getenv("ENGRAM_POOL_FREE_FLOOR_MB");
+    if (fe && *fe){ unsigned long v = strtoul(fe, NULL, 10); if (v) floor_b = (uint64_t)v * 1024ull * 1024ull; }
+    if (avail >= floor_b) return 0;                 /* machine has room */
+    if (!c->cap || c->count == 0) return 0;
+    size_t was = c->cap;
+    size_t want = c->count - (c->count / 4);        /* give back ~25% of what we hold */
+    if (want < 4096) want = 4096;
+    if (want >= c->cap) return 0;
+    c->cap = want;
+    c->adapt_shrinks++;
+    fprintf(stderr,
+        "[engram] memory pressure: %.2f GiB available (floor %.2f GiB) — shrinking pool "
+        "budget %zu -> %zu frames (%.2f -> %.2f GiB) and releasing frames.\n",
+        (double)avail/(1024.0*1024.0*1024.0), (double)floor_b/(1024.0*1024.0*1024.0),
+        was, c->cap,
+        (double)was   * (double)STORE_PAGE_SIZE/(1024.0*1024.0*1024.0),
+        (double)c->cap* (double)STORE_PAGE_SIZE/(1024.0*1024.0*1024.0));
+    fflush(stderr);
+    return 1;
+}
+
+static void pc_adapt_budget(PgCache* c){
+    if (!c->cap) return;                     /* unlimited: nothing to adapt */
+    if (getenv("ENGRAM_POOL_FRAMES")) return; /* explicit operator override wins */
+
+    /* Sliding window so the signal reflects NOW, not lifetime totals. */
+    uint64_t acc = c->hits + c->misses;
+    if (acc - c->adapt_last_acc < 100000) return;
+    uint64_t d_acc  = acc - c->adapt_last_acc;
+    uint64_t d_evic = c->evictions - c->adapt_last_evic;
+    uint64_t d_hits = c->hits - c->adapt_last_hits;
+    c->adapt_last_acc = acc; c->adapt_last_evic = c->evictions; c->adapt_last_hits = c->hits;
+
+    /* Pressure = evicting on a large fraction of accesses while still getting
+     * real reuse. Evictions alone are normal (a scan evicts and never returns);
+     * evictions WITH reuse means the working set genuinely does not fit. */
+    if (d_evic * 3 < d_acc) return;          /* < 1/3 of accesses evict: healthy */
+    if (d_hits * 4 < d_acc) return;          /* little reuse: a scan, not pressure */
+
+    /* Growth is bounded by what is AVAILABLE, never by total RAM. Sizing against
+     * total is how a cache starves its own host: total never shrinks when other
+     * processes need memory. Refuse to grow at all if availability is unknown or
+     * already under the floor — a cache is never worth swapping the machine. */
+    uint64_t avail = pc_available_ram();
+    uint64_t floor_b = ENGRAM_POOL_FREE_FLOOR_BYTES;
+    const char* fe = getenv("ENGRAM_POOL_FREE_FLOOR_MB");
+    if (fe && *fe){ unsigned long v = strtoul(fe, NULL, 10); if (v) floor_b = (uint64_t)v * 1024ull * 1024ull; }
+    if (!avail || avail <= floor_b) return;
+    uint64_t ram = pc_ram_bytes_live();
+    if (!ram) return;
+    unsigned pct = 50;                        /* ceiling as a share of TOTAL, belt-and-braces */
+    const char* mp = getenv("ENGRAM_POOL_MAX_PCT");
+    if (mp && *mp){ unsigned long v = strtoul(mp, NULL, 10); if (v > 0 && v <= 95) pct = (unsigned)v; }
+    size_t ceiling = (size_t)(((ram / 100u) * pct) / (uint64_t)STORE_PAGE_SIZE);
+    /* and never grow into the free-memory floor */
+    uint64_t headroom = avail - floor_b;
+    size_t   ceil_avail = (size_t)((c->count * (uint64_t)STORE_PAGE_SIZE + headroom)
+                                   / (uint64_t)STORE_PAGE_SIZE);
+    if (ceil_avail < ceiling) ceiling = ceil_avail;
+    if (c->cap >= ceiling) return;            /* already at the machine's limit */
+
+    size_t want = c->cap + (c->cap / 2) + 1;  /* ×1.5, geometric */
+    if (want > ceiling) want = ceiling;
+    size_t was = c->cap;
+    c->cap = want;
+    c->adapt_grows++;
+    /* Emit the sensed state, not just the reaction. These are the numbers that
+     * would have diagnosed 2026-08-15 in seconds instead of hours. */
+    pc_report(c, "budget-grow");
+    fprintf(stderr,
+        "[engram] pool pressure: %llu evictions / %llu accesses (%llu hits) at %zu frames "
+        "(%.2f GiB) — working set exceeds budget; growing to %zu frames (%.2f GiB).\n",
+        (unsigned long long)d_evic, (unsigned long long)d_acc, (unsigned long long)d_hits,
+        was,  (double)was  * (double)STORE_PAGE_SIZE / (1024.0*1024.0*1024.0),
+        c->cap,(double)c->cap * (double)STORE_PAGE_SIZE / (1024.0*1024.0*1024.0));
+    fflush(stderr);
+}
+
 static void pc_evict_to_budget(PgCache* c){
     if (!c->cap) return;                    /* unlimited */
     while (c->count > c->cap){
@@ -1687,6 +2042,7 @@ static void pc_evict_to_budget(PgCache* c){
         }
         if (!freed) break;                  /* nothing evictable — allowed to exceed cap */
     }
+    if (!pc_relieve_pressure(c)) pc_adapt_budget(c);
 }
 
 static PgEnt* pc_get(EngramPagedStore* s, uint64_t id){
@@ -2377,6 +2733,18 @@ int store_put_node(EngramPagedStore* s, const StoreNode* n){
 int store_put_edge(EngramPagedStore* s, const StoreEdge* e){
     if (!s || !e || !e->id || !e->from_id || !e->to_id) return -1;
     STORE_GUARD(s);
+    /* Durable-hash write barrier — mirrors store_put_node. An unchanged edge
+     * costs one hash compare and zero page I/O; without this, checkpointing
+     * re-probed every edge against the paged store (max_page_lsn_for_id →
+     * page_read), turning a routine checkpoint into a full-store walk. */
+    uint64_t dh_h = 0;
+    if (s->barrier_on){
+        dh_h = dh_edge_hash(e);
+        if (dh_get(s->dh, e->id) == dh_h){
+            s->stat_barrier_skips++;
+            return 0;
+        }
+    }
     uint64_t L = ++s->next_lsn;
     if (s->wal){
         size_t blen; uint8_t* body = edge_serialize(e, &blen);
@@ -2386,6 +2754,10 @@ int store_put_edge(EngramPagedStore* s, const StoreEdge* e){
         if (wr != 0) return -1;
     }
     int r = apply_edge_put(s, e, L);
+    if (r == 0 && s->barrier_on){
+        if (!dh_h) dh_h = dh_edge_hash(e);
+        dh_set(s->dh, e->id, dh_h);   /* remember the now-persisted durable hash */
+    }
     ckpt_maybe(s);
     return r;
 }

lang/runtime/engram_vindex.c

@@ -74,11 +74,6 @@ struct VIndex {
 
     int      entry;         /* entry-point element index, -1 if empty */
     int      max_level;     /* current top layer */
-
-    /* scratch: version-stamped visited set (O(1) reset). */
-    uint32_t* visited;
-    uint32_t  visit_epoch;
-    size_t    visited_cap;
 };
 
 /* ── small helpers ────────────────────────────────────────────────────────── */
@@ -166,37 +161,63 @@ static Pair heap_pop(Heap* h, int is_max){
     return top;
 }
 
-/* ── visited set ──────────────────────────────────────────────────────────── */
-static int visited_ensure(VIndex* ix){
-    if (ix->visited_cap >= ix->cap && ix->visited) return 0;
-    size_t nc = ix->cap ? ix->cap : 16;
-    uint32_t* nv = (uint32_t*)realloc(ix->visited, nc*sizeof(uint32_t));
-    if (!nv) return -1;
-    if (nc > ix->visited_cap) memset(nv + ix->visited_cap, 0, (nc-ix->visited_cap)*sizeof(uint32_t));
-    ix->visited = nv; ix->visited_cap = nc;
+/* ── visited set — owned by the CALL FRAME, never by the index ──────────────
+ * This buffer is per-TRAVERSAL scratch. It used to live in struct VIndex as an
+ * allocation optimisation, which made every traversal a write to shared state:
+ * two concurrent vindex_search calls stamped each other's epoch and then walked
+ * each other's marks, so even two pure READS corrupted the traversal (measured
+ * 2026-08-16: TSan data race at visited_reset, reached from vindex_search on one
+ * thread and vindex_insert on another; downstream SIGSEGV dereferencing a bogus
+ * element index).
+ *
+ * It is not an ownership problem and it does not want a lock or a capability —
+ * it was simply misfiled. A pure function's scratch belongs to the call. Moving
+ * it here is what lets vindex_search take a `const VIndex*`, which is in turn
+ * what makes "search does not mutate the index" a COMPILE-TIME property instead
+ * of a review comment.
+ *
+ * Cost: one calloc/free of cap*4 bytes per traversal (~55 KB at the live store's
+ * 13,820 elements), against thousands of dim-768 dot products in the same call.
+ * Deliberately NOT __thread: http_worker is a thread per connection, so a
+ * thread-local buffer would retain ~55 KB per connection for the process life. */
+typedef struct {
+    uint32_t* mark;      /* per-element epoch stamp */
+    uint32_t  epoch;     /* current traversal's stamp; 0 == "no traversal yet" */
+    size_t    cap;
+} VVisit;
+
+/* calloc leaves every stamp 0 and epoch 0; the first visit_reset moves to
+ * epoch 1, so no element reads as visited before it is marked. */
+static int visit_init(VVisit* v, size_t cap){
+    size_t nc = cap ? cap : 16;
+    v->mark = (uint32_t*)calloc(nc, sizeof(uint32_t));
+    if (!v->mark) return -1;
+    v->cap = nc; v->epoch = 0;
     return 0;
 }
-static inline void visited_reset(VIndex* ix){
-    if (++ix->visit_epoch == 0){                 /* wrapped: clear all */
-        memset(ix->visited, 0, ix->visited_cap*sizeof(uint32_t));
-        ix->visit_epoch = 1;
+static void visit_dispose(VVisit* v){ free(v->mark); v->mark = NULL; v->cap = 0; }
+static inline void visit_reset(VVisit* v){
+    if (++v->epoch == 0){                        /* wrapped: clear all */
+        memset(v->mark, 0, v->cap*sizeof(uint32_t));
+        v->epoch = 1;
     }
 }
-static inline int is_visited(VIndex* ix, int e){ return ix->visited[e]==ix->visit_epoch; }
-static inline void mark_visited(VIndex* ix, int e){ ix->visited[e]=ix->visit_epoch; }
+static inline int is_visited(const VVisit* v, int e){ return v->mark[e]==v->epoch; }
+static inline void mark_visited(VVisit* v, int e){ v->mark[e]=v->epoch; }
 
 /* ── search one layer (Algorithm 2): best-first, ef-bounded ───────────────── */
 /* Returns results as an unsorted Heap (max-heap on distance, size<=ef). Caller
  * owns res->a. `q` is a normalised query. */
-static int search_layer(VIndex* ix, const float* q, const int* eps, int neps,
+static int search_layer(const VIndex* ix, VVisit* vis, const float* q,
+                        const int* eps, int neps,
                         int ef, int layer, Heap* res /*out, max-heap*/){
     Heap cand = {0,0,0};                          /* min-heap: nearest to expand */
     res->a=NULL; res->n=0; res->cap=0;
-    visited_reset(ix);
+    visit_reset(vis);
     for (int i=0;i<neps;i++){
         int e = eps[i];
-        if (is_visited(ix,e)) continue;
-        mark_visited(ix,e);
+        if (is_visited(vis,e)) continue;
+        mark_visited(vis,e);
         float d = vdist(ix, q, ix->elems[e].vec);
         Pair p = { d, e };
         if (heap_push(&cand,p,0) || heap_push(res,p,1)){ free(cand.a); return -1; }
@@ -212,8 +233,8 @@ static int search_layer(VIndex* ix, const float* q, const int* eps, int neps,
             NeighList* nl = &ce->links[layer];
             for (int i=0;i<nl->count;i++){
                 int e = nl->ids[i];
-                if (is_visited(ix,e)) continue;
-                mark_visited(ix,e);
+                if (is_visited(vis,e)) continue;
+                mark_visited(vis,e);
                 float d = vdist(ix, q, ix->elems[e].vec);
                 if (res->n < ef || d < res->a[0].d){
                     Pair p = { d, e };
@@ -232,7 +253,7 @@ static int search_layer(VIndex* ix, const float* q, const int* eps, int neps,
  * Keep c only if it is nearer to q than to every already-chosen neighbour;
  * backfill from the pruned set (nearest first) to reach M for connectivity.
  * Writes chosen element indices into out[], returns the count. */
-static int select_neighbors(VIndex* ix, const float* q, Pair* W, int nW, int M, int* out){
+static int select_neighbors(const VIndex* ix, const float* q, Pair* W, int nW, int M, int* out){
     (void)q;   /* q's distances are precomputed in W[].d; kept for call-site clarity */
     /* sort W ascending by (dist,elem) — deterministic. */
     for (int i=1;i<nW;i++){                        /* insertion sort (nW small) */
@@ -281,7 +302,7 @@ static int elems_reserve(VIndex* ix){
     Elem* ne = (Elem*)realloc(ix->elems, nc*sizeof(Elem));
     if (!ne) return -1;
     ix->elems = ne; ix->cap = nc;
-    return visited_ensure(ix);
+    return 0;
 }
 
 int vindex_insert(VIndex* ix, uint64_t node_id, const float* vec){
@@ -307,13 +328,19 @@ int vindex_insert(VIndex* ix, uint64_t node_id, const float* vec){
         return 0;
     }
 
+    /* This call frame owns its traversal scratch for the whole insert. ix->cap
+     * already covers `cur` (elems_reserve ran above), so every reachable element
+     * index is in range. */
+    VVisit vis;
+    if (visit_init(&vis, ix->cap)) return -1;
+
     int ep = ix->entry;
     int L = ix->max_level;
     /* greedy descent through layers above `level` to refine the entry point. */
     for (int lc = L; lc > level; lc--){
         Heap r = {0,0,0};
         int eps1[1] = { ep };
-        if (search_layer(ix, el->vec, eps1, 1, 1, lc, &r)){ return -1; }
+        if (search_layer(ix, &vis, el->vec, eps1, 1, 1, lc, &r)){ visit_dispose(&vis); return -1; }
         if (r.n){ ep = r.a[0].e; float bd=r.a[0].d;
             for (int i=1;i<r.n;i++) if (r.a[i].d<bd){bd=r.a[i].d; ep=r.a[i].e;} }
         free(r.a);
@@ -329,7 +356,7 @@ int vindex_insert(VIndex* ix, uint64_t node_id, const float* vec){
     for (int lc = start; lc >= 0; lc--){
         int Mmax = (lc==0) ? ix->M0 : ix->M;
         Heap W = {0,0,0};
-        if (search_layer(ix, el->vec, eps, neps, ix->ef_construction, lc, &W)){ rc=-1; break; }
+        if (search_layer(ix, &vis, el->vec, eps, neps, ix->ef_construction, lc, &W)){ rc=-1; break; }
         int* chosen = (int*)malloc((size_t)(W.n?W.n:1)*sizeof(int));
         if (!chosen){ free(W.a); rc=-1; break; }
         int nc = select_neighbors(ix, el->vec, W.a, W.n, Mmax, chosen);
@@ -357,13 +384,17 @@ int vindex_insert(VIndex* ix, uint64_t node_id, const float* vec){
     }
 done:
     free(eps_owned);
+    visit_dispose(&vis);
     if (rc) return -1;
     if (level > ix->max_level){ ix->max_level = level; ix->entry = cur; }
     return 0;
 }
 
 /* ── search ───────────────────────────────────────────────────────────────── */
-int vindex_search(VIndex* ix, const float* query, int k, int ef_search,
+/* `ix` is const: search is pure with respect to the index. That is enforced by
+ * the compiler, not by convention — it is the whole point of moving the visited
+ * set into the frame below. */
+int vindex_search(const VIndex* ix, const float* query, int k, int ef_search,
                   uint64_t* node_id_out, float* dist_out){
     if (!ix || !query || k <= 0) return -1;
     if (ix->entry < 0) return 0;
@@ -373,11 +404,15 @@ int vindex_search(VIndex* ix, const float* query, int k, int ef_search,
     float* q = vec_normalise_copy(query, ix->dim);
     if (!q) return -1;
 
+    /* This call frame owns its traversal scratch. */
+    VVisit vis;
+    if (visit_init(&vis, ix->cap)){ free(q); return -1; }
+
     int ep = ix->entry;
     for (int lc = ix->max_level; lc > 0; lc--){
         Heap r = {0,0,0};
         int eps[1] = { ep };
-        if (search_layer(ix, q, eps, 1, 1, lc, &r)){ free(q); return -1; }
+        if (search_layer(ix, &vis, q, eps, 1, 1, lc, &r)){ visit_dispose(&vis); free(q); return -1; }
         if (r.n){ int b=r.a[0].e; float bd=r.a[0].d;
             for (int i=1;i<r.n;i++) if (r.a[i].d<bd){bd=r.a[i].d; b=r.a[i].e;}
             ep = b; }
@@ -385,7 +420,8 @@ int vindex_search(VIndex* ix, const float* query, int k, int ef_search,
     }
     Heap res = {0,0,0};
     int eps[1] = { ep };
-    if (search_layer(ix, q, eps, 1, ef_search, 0, &res)){ free(res.a); free(q); return -1; }
+    if (search_layer(ix, &vis, q, eps, 1, ef_search, 0, &res)){ visit_dispose(&vis); free(res.a); free(q); return -1; }
+    visit_dispose(&vis);
     free(q);
 
     /* res is a max-heap of size<=ef; pop into ascending order, keep nearest k. */
@@ -419,7 +455,6 @@ VIndex* vindex_create(int dim, int M, int ef_construction){
     ix->mL = 1.0 / log((double)M > 1.0 ? (double)M : 2.0);
     ix->entry = -1;
     ix->max_level = 0;
-    ix->visit_epoch = 0;
     return ix;
 }
 
@@ -432,7 +467,6 @@ void vindex_free(VIndex* ix){
         free(e->vec);
     }
     free(ix->elems);
-    free(ix->visited);
     free(ix);
 }
 

lang/runtime/engram_vindex.h

@@ -53,8 +53,15 @@ int vindex_insert(VIndex* idx, uint64_t node_id, const float* vec);
  * first (ascending distance). Either out array may be NULL to skip it.
  *   ef_search — search-time candidate width; larger == higher recall, slower.
  *               Pass <=0 for VINDEX_DEFAULT_EF_SEARCH. Internally clamped to >=k.
- * Returns the number of results written, or <0 on error. */
-int vindex_search(VIndex* idx, const float* query, int k, int ef_search,
+ * Returns the number of results written, or <0 on error.
+ *
+ * `idx` is const BY CONTRACT AND BY TYPE: search does not mutate the index. The
+ * traversal's visited set is owned by the call frame, so N threads may search one
+ * index concurrently. Concurrent search against a vindex_insert on the same index
+ * is still unsafe — insert rewires existing elements' neighbour lists and reallocs
+ * elems[] — so the index's owner must not extend a published index under a live
+ * reader. See eg_vindex_view / eg_vindex_maintain in el_runtime.c. */
+int vindex_search(const VIndex* idx, const float* query, int k, int ef_search,
                   uint64_t* node_id_out, float* dist_out);
 
 /* Number of vectors currently indexed. */

lang/runtime/vindex_bench.c

@@ -7,11 +7,29 @@
  *
  * Read-only: never opens a socket, never writes the store. Safe on an nsbx clone.
  *
- * Build:  cc -O2 -std=c11 vindex_bench.c engram_vindex.c -lm -o vindex_bench
+ * Also runs the brute-force oracle a second (and third) way, through the
+ * batch-cosine Strategies behind eg_cosine_batch_strategy.h — the ggml
+ * strategy and the hand-rolled-Metal strategy (Apple/Metal only; see
+ * eg_cosine_batch.h/eg_cosine_batch_strategy.h) — and reports each one's
+ * latency + a correctness check against the CPU oracle side-by-side with the
+ * existing CPU-vs-HNSW numbers. This harness deliberately reaches past the
+ * single-selection Factory (eg_cosine_batch.c) to instantiate every
+ * compiled-in strategy directly, so it can compare all of them against the
+ * SAME dataset in one run — that is the harness's whole job; a real call
+ * site (el_runtime.c) never does this, it only ever calls the plain
+ * eg_cosine_batch()/eg_cosine_batch_multi() adapter functions.
+ * EL_METAL_COSINE=0 forces CPU-only (skips every strategy comparison).
+ *
+ * Build (macOS, ggml + hand-rolled Metal): see build_vindex_bench.sh.
+ * Build (Linux / no Metal): omit every eg_cosine_batch_strategy_*.{c,m} file
+ *   except eg_cosine_batch_strategy_cpu.c — this file never references
+ *   ggml/Metal directly except through the plain-C strategy header, guarded
+ *   by the same EG_HAVE_STRATEGY_* build macros the Factory itself uses.
  * Usage:  vindex_bench store <neuron.egm> <dim> [nqueries] [k] [ef_csv]
  *         vindex_bench synth <N> [dim] [clusters] [nqueries] [k] [ef_csv]
  */
 #include "engram_vindex.h"
+#include "eg_cosine_batch_strategy.h"
 #include <stdio.h>
 #include <stdlib.h>
 #include <string.h>
@@ -74,6 +92,106 @@ static double recall_at_k(const int* gt, const uint64_t* ann, int nann, int k){
     return (double)hit / (double)k;
 }
 
+/* EL_METAL_COSINE: 0/off/false disables EVERY strategy comparison outright
+ * (falls back to brute_topk() only), matching el_runtime.c's own gate for
+ * the same env var (back-compat name kept from PR #114; it now gates all
+ * GPU-backed strategies, not just the hand-rolled Metal one). Unset or any
+ * other value = try every compiled-in strategy, report each that's
+ * available, skip (without failing the run) any that isn't. */
+static bool g_strategy_env_checked = false;
+static bool g_strategy_disabled_by_env = false;
+static void eg_strategy_check_env_once(void){
+    if (g_strategy_env_checked) return;
+    g_strategy_env_checked = true;
+    const char* v = getenv("EL_METAL_COSINE");
+    if (v && (v[0]=='0' || v[0]=='n' || v[0]=='N' || v[0]=='f' || v[0]=='F'))
+        g_strategy_disabled_by_env = true;
+}
+
+/* Batched sibling of brute_topk, generalized over ANY EgCosineBatchStrategy:
+ * computes top-k for ALL nq queries in ONE strategy->batch_multi() call,
+ * uploading/preparing the node population exactly once instead of once per
+ * query. out_ids/out_d are nq*k, row-major (query i's results at
+ * out_ids+i*k / out_d+i*k). Returns false (nothing written) on any
+ * failure/unavailability; caller treats that as "skip this strategy in the
+ * report", never as a hard error. */
+static bool batch_topk_strategy(const EgCosineBatchStrategy* strat,
+                                const float* data, int n, int dim,
+                                const float* queries, int nq,
+                                int k, int* out_ids, float* out_d){
+    if (!strat || !strat->available()) return false;
+
+    const float** row_ptrs = malloc((size_t)n * sizeof(float*));
+    int32_t* dims = malloc((size_t)n * sizeof(int32_t));
+    double* scores = malloc((size_t)nq * (size_t)n * sizeof(double));
+    if (!row_ptrs || !dims || !scores) { free(row_ptrs); free(dims); free(scores); return false; }
+
+    for (int i = 0; i < n; i++) { row_ptrs[i] = data + (size_t)i * dim; dims[i] = dim; }
+
+    bool ok = strat->batch_multi(queries, dim, nq, row_ptrs, dims, n, scores);
+    free(row_ptrs); free(dims);
+    if (!ok) { free(scores); return false; }
+
+    for (int qi = 0; qi < nq; qi++) {
+        int* ids = out_ids + (size_t)qi * k;
+        float* ds = out_d + (size_t)qi * k;
+        const double* srow = scores + (size_t)qi * n;
+        for (int i = 0; i < k; i++) { ids[i] = -1; ds[i] = 3.0f; }
+        for (int i = 0; i < n; i++) {
+            float d = 1.0f - (float)srow[i];   /* same distance convention as brute_topk */
+            if (d >= ds[k-1]) continue;
+            int p = k - 1;
+            while (p > 0 && ds[p-1] > d) { ds[p] = ds[p-1]; ids[p] = ids[p-1]; p--; }
+            ds[p] = d; ids[p] = i;
+        }
+    }
+    free(scores);
+    return true;
+}
+
+/* Runs batch_topk_strategy for one named strategy over ALL nq queries, diffs
+ * against the CPU ground truth (gt/gd, both nq*k), and prints a report line
+ * in the same shape PR #114 established for BRUTE-METAL — id-recall over
+ * every query plus the actual max/mean same-rank distance delta across
+ * every (query,rank) pair that was compared, never fabricated or assumed. */
+static void report_strategy_vs_oracle(const char* label, const EgCosineBatchStrategy* strat,
+                                      const float* data, int n, int dim,
+                                      const float* qv, int nq, int k,
+                                      const int* gt, const float* gd, double brute_ms){
+    if (g_strategy_disabled_by_env) { printf("%-13s: disabled via EL_METAL_COSINE\n", label); return; }
+    if (!strat || !strat->available()) { printf("%-13s: not available on this build/host — skipped\n", label); return; }
+
+    int* gtm = malloc((size_t)nq*k*sizeof(int));
+    float* gdm = malloc((size_t)nq*k*sizeof(float));
+    double tm0 = now_s();
+    bool ok = batch_topk_strategy(strat, data, n, dim, qv, nq, k, gtm, gdm);
+    double strat_ms = (now_s()-tm0)*1000.0/nq;
+    if (ok) {
+        double rec_sum = 0; double max_ddiff = 0; double sum_ddiff = 0; int compared = 0;
+        for (int i=0;i<nq;i++) {
+            const int* ids_gt = gt+(size_t)i*k;
+            const float* d_gt = gd+(size_t)i*k;
+            const int* ids_m = gtm+(size_t)i*k;
+            const float* d_m = gdm+(size_t)i*k;
+            uint64_t idset[512]; int m = (k<512)?k:512;
+            for (int j=0;j<m;j++) idset[j] = (uint64_t)ids_m[j];
+            rec_sum += recall_at_k(ids_gt, idset, m, k);
+            for (int j=0;j<k;j++) {
+                if (ids_gt[j] == ids_m[j]) {
+                    double diff = fabs((double)d_gt[j]-(double)d_m[j]);
+                    if (diff>max_ddiff) max_ddiff=diff;
+                    sum_ddiff += diff; compared++;
+                }
+            }
+        }
+        printf("%-13s: %8.3f ms/query   (%.1fx vs CPU brute; id-recall %.4f vs CPU oracle over %d queries; same-rank |Δdist|: max %.2e, mean %.2e over %d compared)\n",
+               label, strat_ms, brute_ms/strat_ms, rec_sum/nq, nq, max_ddiff, compared?sum_ddiff/compared:0.0, compared);
+    } else {
+        printf("%-13s: batch call failed mid-run — skipped\n", label);
+    }
+    free(gtm); free(gdm);
+}
+
 /* Parse "64,128,256" into an int array; returns count. */
 static int parse_csv(const char* s, int* out, int maxo){
     int n=0; if(!s||!*s) return 0;
@@ -142,13 +260,37 @@ static void run_bench(const char* label, float* data, int n, int dim,
         l2norm(dst, dim);
     }
 
-    /* ground truth: brute-force top-k for every query (also the oracle latency). */
+    /* ground truth: brute-force top-k for every query (also the oracle latency).
+     * gd is nq*k (one real slot per query, not a shared scratch buffer) so the
+     * strategy comparisons below can diff against every query's actual
+     * distances, not just whichever query happened to run last. */
     int* gt = malloc((size_t)nq*k*sizeof(int));
-    float* gd = malloc((size_t)k*sizeof(float));
+    float* gd = malloc((size_t)nq*k*sizeof(float));
     double tb0 = now_s();
-    for (int i=0;i<nq;i++) brute_topk(data, n, dim, qv+(size_t)i*dim, k, gt+(size_t)i*k, gd);
+    for (int i=0;i<nq;i++) brute_topk(data, n, dim, qv+(size_t)i*dim, k, gt+(size_t)i*k, gd+(size_t)i*k);
     double brute_ms = (now_s()-tb0)*1000.0/nq;
-    printf("BRUTE-FORCE  : %8.3f ms/query   (oracle; O(N*D))\n", brute_ms);
+    printf("BRUTE-FORCE  : %8.3f ms/query   (oracle; O(N*D), CPU)\n", brute_ms);
+
+    /* GPU-backed oracles: SAME nq queries, SAME top-k contract, via each
+     * compiled-in Strategy's batch_multi() (uploads/prepares the node
+     * population once, not once per query). Run only for strategies that
+     * are actually available (checked internally) — never fabricated, never
+     * assumed. Verified against the CPU ground truth computed above:
+     * id-recall across ALL nq queries, plus the actual max/mean distance
+     * delta across every (query,rank) pair that was compared. */
+    eg_strategy_check_env_once();
+#ifdef EG_HAVE_STRATEGY_GGML
+    report_strategy_vs_oracle("BRUTE-GGML", eg_cosine_batch_strategy_ggml(),
+                              data, n, dim, qv, nq, k, gt, gd, brute_ms);
+#else
+    printf("BRUTE-GGML   : strategy not compiled into this build\n");
+#endif
+#ifdef EG_HAVE_STRATEGY_METAL_HAND
+    report_strategy_vs_oracle("BRUTE-METAL", eg_cosine_batch_strategy_metal_hand(),
+                              data, n, dim, qv, nq, k, gt, gd, brute_ms);
+#else
+    printf("BRUTE-METAL  : strategy not compiled into this build\n");
+#endif
 
     /* HNSW at each ef. */
     uint64_t* aid = malloc((size_t)k*sizeof(uint64_t));

lang/spec/runtime-ownership.md

@@ -0,0 +1,180 @@
+# El Runtime — Ownership and Capability ABI
+
+**Status:** §0–§2 verified. §3 re-derived and **built** for the vector index (2026-08-16); not yet applied to the resident RAM graph.
+**Date:** 2026-08-16
+**Scope:** `lang/runtime/` — every El program (soul, engram, cgi-studio vessels) inherits this by rebuild. Nothing in this document is a change to any El *program*.
+
+**Note on §1's line numbers:** they were read against a checkout that has since shifted by ~135 lines. Verified positions as of `a67452f` are in §2a.
+
+---
+
+## 0. The residual
+
+> **Builtins own memory and reach process state directly.**
+
+That is the residual — the generator. Everything below labelled a "residue" is a deposit left by it. The distinction matters because we have spent significant effort removing deposits, and deposits regenerate.
+
+A residue is fixed. A residual is eliminated. Fixing residues while the residual stands produces exactly the pattern observed on 2026-08-15/16: a run of individually-correct patches, each verified, followed by a new defect of the same shape in a different file.
+
+---
+
+## 1. The residues, measured
+
+Each of these is a distinct merged or proposed fix. Each addresses one deposit. None addresses the residual.
+
+| residue | location | fix that was applied or proposed |
+|---|---|---|
+| `state_get` leaked its return value per call — 15 MB over 200k calls | builtin | el #140 (merged) |
+| VIndex freed under a concurrent reader | `el_runtime.c:9424` | `fb32d15` guard (merged 08:46:43) |
+| `_eg_vindex_seen` realloc'd on a read path | `el_runtime.c:9412` | same guard |
+| `vindex_insert` on a read path | `el_runtime.c:9434`, `9450` | same guard |
+| shared `visited` / epoch scratch stomped by concurrent searches | `engram_vindex.c:79–81`, `169–186`, `195` | proposed: move to per-search frame |
+| nine append sites, none indexing → lazily-embedded nodes invisible | `el_runtime.c:7806, 7988, 8148, 8224, 11526, 11731, 12050, 15295, 15312` | "embed-gap #20", patched by making the *read* path catch up (`9439` comment) |
+
+**Measured:** all file/line references above, read 2026-08-16. Crash frames `engram_activate → eg_vindex_sync → vindex_insert → _realloc → _xzm_xzone_malloc_freelist_outlined` are accounted for by rows 2–4.
+
+**Inferred, not yet verified:** that the nine append sites do not share a single commit point. This needs one pass before Change C is sized.
+
+---
+
+## 2. Why these are one defect
+
+`eg_vindex_sync` (`el_runtime.c:9419`) has exactly three callers, and **all three are reads**:
+
+- `engram_activate` — `9802`
+- `eg_knn_for_node` — `13075` (its own header comment states *"No writes."*)
+- `engram_geo_reify_run_json` — `13285`
+
+It mutates five process-global statics (`9400–9404`): `_eg_vindex`, `_eg_vindex_dim`, `_eg_vindex_built_nc`, `_eg_vindex_seen`, `_eg_vindex_seen_cap`.
+
+Reads mutate because index maintenance was never given an owner on the write side. It got bolted onto reads, because a builtin *could* reach the globals — nothing prevented it. Likewise `state_get` leaked because a builtin *owned* the value it returned; nothing prevented that either.
+
+The store is architecturally append-only and superseding. A read path that mutates contradicts that directly. The contradiction is expressible only because the ABI permits it.
+
+---
+
+## 2a. Verified positions and the fact §1 missed
+
+Read directly at `a67452f`, 2026-08-16. §1's line numbers predate a ~135-line shift; these are current.
+
+| thing | §1 said | actually |
+|---|---|---|
+| five process-global statics | 9400–9404 | **9535–9539** |
+| `eg_vindex_seen_ensure` realloc | 9412 | **9547** |
+| `eg_vindex_sync` | 9419 | **9554** |
+| `vindex_free` on a read path | 9424 | **9559** |
+| `vindex_insert` on a read path | 9434 / 9450 | **9569** (build) / **9585** (incremental) |
+| caller: `engram_activate_inner` | 9802 | **9939** |
+| caller: `eg_knn_for_node` | 13075 | **13212** |
+| caller: `engram_geo_reify_run_json` | 13285 | **13422** |
+| `fb32d15` guard | — | lock **1602**, depth **1631**, `eg_guard_enter` **1636**, `http_worker` acquire **1687**, `engram_activate` wrapper **14097** |
+| VIndex scratch fields | 79–81 | **79–81** ✓ |
+| `search_layer` race site | 195 | **195** ✓ |
+
+**The structural fact §1 and §3 both missed:** *the index does not inherit the store's append-only property.* `vindex_insert` rewires the `NeighList` links of already-existing elements and reallocs `elems[]` — so extending the index mutates the whole structure, not just its tail. This is why "make reads pure" is necessary but **not sufficient**, and why §3 needed a publication boundary rather than only a capability split. It is reproduced as a standing test (`unsynchronized` half, §5).
+
+---
+
+## 3. The change
+
+*(Re-derived 2026-08-16. The previous §3 — a runtime context struct carrying read/write **capability pointers** to every builtin — was written in mutable-store, C-ownership terms. It asked "who is permitted to mutate the shared thing?", which presupposes a shared mutable thing. The engram is immutable and recall is projection; what does not mutate needs no ownership discipline. So the question is not answered, it is dissolved. The implemented change is below.)*
+
+### 3.1 Three moves, in decreasing order of how much they dissolve
+
+**(1) Misfiled scratch is not shared state.** `visited` / `visit_epoch` were never conceptually owned by the index — they are one traversal's local, hoisted into `struct VIndex` as an allocation optimisation. Nothing about them is derived geometry. They want neither a lock nor a capability nor a checkout pool: a pure function's scratch belongs to its call frame, and the fix is to put it back there. This is not "the capability model applied by hand to one global"; it is the deletion of a false ownership claim.
+
+**(2) `const` is the capability, and immutability hands it over for free.** Once the scratch leaves the struct, `search_layer` reads the index and nothing else — so `vindex_search` can take a `const VIndex*`. That is *precisely* the teeth old-§3 wanted from capability pointers: a read path physically cannot call `vindex_insert`, and it is a **compile error**, not a review comment. It costs one qualifier rather than a new ABI swept across hundreds of builtins. The compiler enforces it on every future caller for the same reason.
+
+> The capability type was already in the language. It is spelled `const`.
+
+**(3) What remains is a publication problem, not an ownership problem.** With scratch in the frame and reads const, one hazard survives, and it is real: **HNSW insert is not an append.** `vindex_insert` rewires the `NeighList` links of *already-existing* elements and reallocs `elems[]`. The store's append-only property does **not** transfer to the index derived from it. So a reader projecting against the index while its owner extends it is unsafe no matter how pure search is.
+
+Immutability answers this too, and the answer is publication:
+
+- **`eg_vindex_maintain`** — the sole mutator. Takes the boundary exclusively; never runs beside a reader.
+- **`eg_vindex_view`** — returns a `const VIndex*` with the boundary held for read. N readers project concurrently; none can mutate.
+
+A read path may **demand that a current snapshot exist** — that is a request to the owner, not a mutation by the reader. What it may not do is mutate the geometry it is projecting against. `view` / `maintain` is exactly that split, and it is why this replaces `eg_vindex_sync` rather than wrapping it.
+
+**Write-side owner.** Index membership is owned by the event *"an embedding became present on this ordinal"* — not by node append, since a node without an embedding cannot be in a vector index at all. `eg_vindex_note_embedded` hooks the embedding-assignment sites: one O(log n) insert, no O(node_count) presence scan. This also retires the "STALENESS (honest tradeoff)" note in the old `eg_vindex_sync`, where a lazily-embedded *older* node stayed invisible to `route_nearest` / autoconnect until the next full rebuild.
+
+### 3.2 What this does not claim
+
+The **resident RAM graph** (`g->nodes` / `g->edges`) is a *separate* residue of the same residual and is untouched by this change. It is realloc'd in place (`el_runtime.c:7618`, `7629`), so an awareness-thread reader holding `EngramNode* n = &g->nodes[i]` across a concurrent append holds a dangling pointer — and `engram_activate_inner`'s embed-backfill writes `n->emb` through exactly such a pointer. It wants the same publication treatment the index just received. Until that lands, the `fb32d15` guard stays (see §5).
+
+---
+
+## 4. Why this is not a large change
+
+The old §4 argued that El owning its compiler makes a capability-ABI sweep mechanical, since `elc` generates every builtin call site. That argument was load-bearing only for the ABI, and the ABI is gone.
+
+The constraint now travels with the **type of the thing**, not the shape of every call site — so no sweep is needed at all. Measured extent of the implemented change: two qualifiers (`const VIndex*` on `vindex_search`, propagated to `engram_geometry_descriptor` and `engram_geo_reify_store`), one struct field group relocated to a call frame, one rwlock, and three read call sites converted from `eg_vindex_sync` to `view`/`release`.
+
+The payoff of owning the language is unchanged and is now *cheaper*: introduced once, enforced by the compiler on every future builtin, cannot subsequently be forgotten. Contrast the current state, where the same discipline was maintained by hand across hundreds of builtins and demonstrably failed at least six times.
+
+---
+
+## 5. What this deletes
+
+**Deleted (done, 2026-08-16):**
+
+- `eg_vindex_sync` — the function itself. Not renamed: split into `eg_vindex_maintain` (mutating, exclusive, sole owner) and `eg_vindex_view` (const, shared). A name that meant "read paths repair the index" had to stop existing.
+- `VIndex::visited` / `visit_epoch` / `visited_cap` — the struct fields, `visited_ensure`, its call from `elems_reserve`, `ix->visit_epoch = 0` in `vindex_create`, and `free(ix->visited)` in `vindex_free`.
+- The **proposed** per-search scratch *struct on the index* (a checkout pool / `VisitedListPool`) — never built. The buffer is a plain frame local; a pool is machinery for an ownership question that no longer exists.
+- The **proposed** reader-view / owner-handle split for VIndex specifically — superseded. `const` already is the reader view.
+- `EXPECT_RACE` in `run_vindex_concurrency_tests.sh` — a knob that let a known defect ride as "expected". Replaced by four halves with real verdicts.
+
+**NOT deleted — the design doc was wrong about this one:**
+
+- `fb32d15` (`eg_guard_enter` / `engram_req_lock` / `_eg_req_depth`). §5 originally called for its removal as "a lock protecting a mutation that ceases to exist." **Measured, it guards two things, and only one of them ceases to exist.** Its own comment names both: the RAM graph *and* `_eg_vindex`. The vindex justification is retired; the RAM-graph justification is independently load-bearing (§3.2), and removing the guard reintroduces the measured 11171→9579 edge-loss defect from 2026-08-14. Its comment has been narrowed to state the RAM graph only. **Precondition for deleting it:** the resident graph gets the same publication boundary the index just got.
+- el #140's hand-patch. Left in place — the leak stops being *expressible* only under the abandoned capability-ABI §3, which is not what was built.
+
+**Ordering consequence (revised):** the original ordering claim — "the residual lands first, the residues evaporate rather than get fixed" — did not survive contact. The residual here is not a single ABI that dissolves everything at once; it is a *property* (derived state is published, never edited) applied per structure. The index now has it. The RAM graph does not yet. Residues evaporate **per structure, in the order the property is applied**, and a residue whose structure has not been converted must be left standing, not deleted on the strength of the plan.
+
+---
+
+## 6. Sequencing
+
+1. **Read** how builtins are declared and dispatched, to confirm the call sites are compiler-generated in one place. *(This determines whether §4 holds. If dispatch is scattered, re-size before proceeding.)*
+2. Introduce the context type and capability types.
+3. Codegen emits the context at every builtin call site.
+4. Mechanical sweep of builtin signatures.
+5. Move index maintenance behind the write capability; the three read callers take the read capability.
+6. Delete the residue-fixes listed in §5.
+7. **One** build of soul from el dev — which resolves the `state_get` leak and the crash together, rather than deploying a leak fix that reintroduces the crash.
+
+---
+
+## 7. Open questions
+
+**Answered 2026-08-16:**
+
+- ~~Do the nine append sites share a commit point?~~ **Moot.** The question was mis-aimed: node append is not the event that owns index membership, because a node without an embedding cannot be in a vector index. The five *embedding-assignment* sites are the real owner points (`el_runtime.c:7091, 9839, 13362, 15002`, plus snapshot-restore at `7951`), and three of them carry the ordinal directly — which is all `eg_vindex_note_embedded` needs. The other two run before the node is resident, where the cold build picks it up.
+- ~~Does anything outside `lang/runtime/` construct a second `VIndex`?~~ **No.** Swept: the only constructors outside the runtime are `engram/test/*` and `lang/runtime/vindex_bench.c`, all single-threaded and index-private. Inside the runtime, `engram_self_reify_beat_json` builds a **private** index deliberately and never touches the shared boundary — that was already correct and is unchanged.
+- ~~Does the HTTP worker pool contend on the same globals?~~ **Yes, and it was never the whole story.** Workers serialize against each other on `engram_req_lock`, but the awareness main thread does not take it at all — that is the gap `fb32d15` closed. Now verified independent of that guard: the index boundary is its own rwlock, so worker/awareness contention on `_eg_vindex` is handled whether or not the request lock is held.
+
+**Still open:**
+
+- The resident RAM graph wants the same publication boundary (§3.2). Until it has one, `fb32d15` cannot be deleted.
+- `eg_vindex_view` holds the boundary for read across `engram_geo_reify_store`, which is a long pass. Correct, but it stalls the owner for that duration. If reify latency becomes a problem the answer is a refcounted snapshot, not a shorter lock.
+
+---
+
+## 7a. Evidence (measured 2026-08-16, `engram/test/run_vindex_concurrency_tests.sh`)
+
+| half | before | after |
+|---|---|---|
+| `single` — 3000 vectors, 1 thread, ASan+UBSan | clean | clean |
+| `readers` — 4 readers, no writer, TSan | **race** at `engram_vindex.c:195` (`visited_reset` ← `vindex_search`) | **clean** |
+| `unsynchronized` — writer+reader, bare index, TSan | race | **race, expected and permanent** — now the proof the boundary must exist |
+| `published` — owner + 4 readers through the boundary, TSan | *(did not exist)* | **clean**, all 3000 inserts landed |
+
+No recall regression: `recall@10 = 0.9365` at `ef_search=128` (gate ≥ 0.90); the determinism test still yields byte-identical results across two independent builds.
+
+Builds locally: all seven engram runtime translation units compile `-Wall -Wextra` clean, and the full engram binary links (`engram/dist/engram.c` + runtime, arm64). The one pre-existing `-Wcomment` warning in `el_runtime.c` is present at `a67452f` too.
+
+---
+
+## 8. What this document is not
+
+It is not an argument for a memory model in general, a garbage collector, process isolation between soul and engram, or a client/server split of the store. Each of those was considered and each addresses mutation that this change removes. They are answers to a question that stops being asked.

lang/swarm/README.md

@@ -0,0 +1,184 @@
+# Swarm + CCR + Work-Tracking — Neuron's bounded parallel execution, in native El
+
+Bounded parallel agent execution on El's **native** concurrency — no external
+orchestrator. Grounded directly in two of Will's frameworks:
+
+- **Swarm Architecture** (*Bounded Parallel Agent Execution*, Mar 2026)
+- **Compiled Context Runtime / CCR** (*Process-Driven Agent Execution with
+  Unbounded Local Memory*, Mar 2026)
+
+A swarm is a **coordinator** (the main thread) that mints a correlation identity,
+compiles a **bounded per-worker context (CCR)**, dispatches workers as **native
+pthreads** (`thread.el` `spawn`/`join`), tracks every unit of work durably, and
+**converges** results before returning control to the parent step.
+
+```
+Parent step
+  └─ swarm_run(blueprint, knowledge_refs, inputs, config)
+        fan-out ──▶ worker_1 (CCR ctx_1) ─┐   native
+                    worker_2 (CCR ctx_2) ─┤   pthreads,
+                    worker_k (CCR ctx_k) ─┘   bounded by `concurrency`
+        converge ─▶ collect | merge | vote | reduce  ──▶ merged result
+```
+
+## Why it runs on El natively
+
+El is natively agentic. This capability composes El's shipped primitives — it
+adds no bespoke runtime:
+
+| Primitive | Source | Role in the swarm |
+|-----------|--------|-------------------|
+| `spawn(fn,arg)` / `join(tid)` | `runtime/thread.el` → `__thread_create` (pthread + dlsym) | fan-out / rejoin |
+| `parallel_map`, `with_mutex` | `runtime/thread.el` | reference concurrency patterns |
+| Go-style channels | `runtime/channel.el` → `__channel_*` | available for vertical event streams |
+| `engram_*`, `http_*`, `fs_*`, `json_*` | `el_runtime.c` builtins | retrieval, tracking, I/O |
+
+Every El fn compiles to a global C symbol, so any top-level `(String)->String`
+fn is directly threadable — the worker entry is exactly such a fn.
+
+## Modules
+
+| File | Framework grounding | What it does |
+|------|--------------------|--------------|
+| `worktrack.el` | Swarm §6 (correlation IDs, audit) | Durable, single-writer **JSONL journal** keyed by correlation ID; reconstructable status report; opt-in engram mirror (`SWARM_MIRROR=1`). |
+| `containment.el` | Swarm §3 + the single-writer invariant | Scope tokens w/ capabilities; **Rule 1** (no join), **Rule 2** (no open), **Rule 3** (no lateral edge), **Rule 4** (engram-write is @manager-only, by capability) enforced as checks. |
+| `ccr.el` | CCR §5 + Swarm §9.3 | Per-worker **Compiled Context Routing**: retrieve → scope → compact into a **bounded, minimal** package. The compiled-context boundary *is* the security boundary. |
+| `primitives.el` | CCR §2 (Five Primitives) | `attend / think / intend / act / learn` seam the swarm composes over. Engram-backed; explicit binding point for the API-surface reshape. |
+| `swarm.el` | Swarm §2, §4, §5 | The coordinator: fan-out/converge on native threads, bounded concurrency, four convergence strategies, integer failure threshold, full tracking. |
+
+## Invariant: only the orchestrator mutates global engram state
+
+**Only the orchestrator (@manager) writes to the engram / mutates global state.
+Workers are read-only against the full engram and may write only their own local
+geometry (their returned result + the journal). A worker is STRUCTURALLY UNABLE
+to mutate global engram state.**
+
+This is **Rule 4** — an **authority gate, not a health gate**. Scope tokens carry
+a capability set: the orchestrator's token holds `engram:write` + `dharma:emit`
+(@manager-only, the VBD rule that only the manager mutates global state); a
+worker's token holds **only** `engram:read`. Every engram mutation
+(`op_write`/`op_relate`/`op_supersede` → `POST /api/nodes`, `/api/edges`,
+`DELETE`) flows through `swarm_engram_write`, which checks the caller's capability
+via the **same scope-token mechanism as the live Rule-2 denial** and rejects any
+worker **before any HTTP is issued**. Capability is fixed at mint time and cannot
+be acquired at runtime — so the guarantee holds regardless of engram health
+(distinct from the `SWARM_WRITE_HEALTHY` *health* gate).
+
+The **curated merge is the only write path**: workers return geometry; the
+orchestrator, and only the orchestrator, commits the approved/verified geometry
+back (`commit=1`). Workers keep full-engram **read** access (`op_think`/`op_read`).
+
+Proven in `harness_real_cognition.el` (§G): a worker `swarm_engram_write` is
+DENIED by capability with no node created and the violation journalled; the
+orchestrator passes the gate as the sole authorized writer.
+
+## Containment → distribution
+
+The three containment rules make workers **location-independent** (Swarm §9): a
+worker reads only its compiled context, shares no state with siblings, and its
+only outward edge is the returned result. The same coordinator can run workers
+as local threads today or dispatch them across machines later — the mechanism is
+identical; only the topology changes. Enforced here:
+
+- **Rule 2** — `swarm_run` rejects any swarm opened under a worker token.
+- **Rules 1 + 3** — each worker gets a *closed* worker token; the coordinator is
+  the only journal writer, so workers share no mutable state.
+
+## Usage
+
+```el
+// one process step fans out; results converge before the next step
+let inputs: String = "[\"billing\",\"payments\",\"ledger\"]"
+let refs:   String = "[\"Volatility-Based Decomposition\"]"   // CCR knowledge refs
+let cfg:    String = "{\"concurrency\":\"4\",\"strategy\":\"collect\",\"min_success_ratio\":\"1.0\"}"
+let result: String = swarm_run("analyze_item", refs, inputs, cfg)
+// result: { corr_id, status, merged, report }
+```
+
+Build any program that uses the swarm:
+
+```bash
+lang/swarm/build.sh myprog.el ./myprog     # concat + elc + cc (el_runtime.c)
+```
+
+Config keys: `concurrency` (max workers at once), `strategy`
+(`collect|merge|vote|reduce`), `min_success_ratio` (decimal string, e.g. `0.8`),
+`caller_token` (containment). Env: `SWARM_TRACK_DIR` (journal dir),
+`CCR_TOKEN_BUDGET`, `ENGRAM_URL`/`ENGRAM_API_KEY` (retrieval + mirror),
+`SWARM_MIRROR=1`.
+
+## Tests
+
+```bash
+lang/swarm/build.sh lang/swarm/tests/test_swarm.el       /tmp/t && SWARM_TRACK_DIR=/tmp/trk /tmp/t   # 12/12
+lang/swarm/build.sh lang/swarm/tests/test_convergence.el /tmp/c && SWARM_TRACK_DIR=/tmp/trk /tmp/c   #  8/8
+# integration against an isolated engram clone (never live):
+source <sandbox>/.nsbx-env
+lang/swarm/build.sh lang/swarm/tests/integ_engram.el /tmp/i && /tmp/i
+```
+
+## Local-swarm integration harness (the one flip)
+
+`tests/harness_local_swarm.el` proves the **full local-swarm mechanics today** on
+the isolated clone with the primitive seam pointed at the hermetic stub — 17/17
+green: 8 native-thread workers at concurrency 4, reduce + vote convergence, CCR
+scoping + non-leak, all three containment rules (incl. live Rule-2 denial),
+durable work-tracking, and **afferent telemetry** observed by the @manager.
+
+Binding to the reshape's decorated primitives is **one flip and a run**:
+
+```
+# in primitive_binding.el — change one line each:
+fn bound_think(ctx, instruction) { return think(ctx, instruction) }   # decorated, dharma bus
+# then:
+SWARM_PRIMITIVE_SEAM=decorated  lang/swarm/build.sh tests/harness_local_swarm.el ./h && ./h
+```
+
+Nothing else in the swarm changes. `primitive_seam.el` (`seam_think/attend/learn`)
+already routes every worker primitive call through this one switch, and the same
+harness runs the bound path. Today `SWARM_PRIMITIVE_SEAM=decorated` still runs
+green because the binding falls back to the stub — proving the flip path executes.
+
+## Real cognition — the seam is BOUND
+
+`primitive_binding.el` is bound to the api-reshape agent's proven primitives
+(`wt/api-reshape@d4f401d`): `bound_think -> op_think` (GET `/api/think`), real
+768-dim gradients over the engram geometry. `reshape_surface.el` composes those
+read/cognition primitives verbatim (`op_think/read/attend/learn`).
+
+`tests/harness_real_cognition.el` runs the **local swarm on real cognition**,
+17/17 green with `SWARM_PRIMITIVE_SEAM=decorated` against the `:8901` clone: 8
+native-thread workers, each a real `think` over its CCR-scoped **node-id anchor**
+(free-text anchors return "geometry unavailable"), `@manager` reduce+vote, all
+three containment rules, afferent telemetry, durable tracking. Per-anchor support
+counts (e.g. 6 / 16 / 87) drive a genuine, cognition-derived vote.
+
+> **Build note (load-bearing):** the swarm build **must** define `HAVE_CURL`
+> (`build.sh` does). Without it every `http_*` builtin is a
+> `{"error":"not built with HAVE_CURL"}` stub — real HTTP silently disappears.
+
+Writes (`attend`/`learn`, `POST`) are gated behind `SWARM_WRITE_HEALTHY=1` and the
+api-reshape agent's gate-1 write-healthy clone; the proven run is read-cognition.
+
+## Built vs stubbed (honest)
+
+**Real, tested:**
+- Native-thread fan-out/converge, bounded concurrency, order-preserving rejoin.
+- All three containment rules enforced (scope tokens + lateral-edge check).
+- CCR per-worker context: retrieval → scoping → compaction, bounded, non-leaking
+  (a worker never receives sibling inputs) — verified against the live isolated mind.
+- Full durable work-tracking (JSONL journal, reconstructable report).
+- Four convergence strategies + integer failure threshold / partial-abort.
+
+**Seam / not yet bound:**
+- `primitives.el` `think` is a deterministic, hermetic transform (no model call).
+  Binding point is marked `PRIMITIVE_BINDING`; wire to the API-surface reshape's
+  `think/act/attend/intend/learn` when it lands.
+- Blueprints are dispatched by name in `swarm_run_blueprint` (default +
+  `classify`/`faildemo` demos). A YAML process-definition loader (Swarm §5) is
+  future work — the runtime contract is in place.
+- Distributed placement (cloud/edge/federated topologies, Swarm §9.2) is
+  structurally enabled by containment but not yet wired to a placement layer;
+  today all workers are local native threads.
+- Engram work-tracking mirror is opt-in; the durable substrate is the journal.
+

lang/swarm/build.sh

@@ -0,0 +1,61 @@
+#!/usr/bin/env bash
+# build.sh — compile an El program that uses the swarm capability.
+#
+# Concatenates the El native-concurrency stdlib (thread.el, channel.el) and the
+# swarm capability modules in dependency order, then the user program, compiles
+# with the canonical elc, and links against the shared C runtime.
+#
+# Usage:
+#   swarm/build.sh <program.el> <out-binary>
+#
+# The swarm modules use only el_runtime.c builtins plus thread.el/channel.el,
+# so nothing else needs concatenating (engram_*, json_*, str_*, fs_*, http_*,
+# uuid_v4, now_millis are all C builtins in el_runtime.c).
+
+set -uo pipefail
+cd "$(dirname "$0")/.."          # -> lang/
+LANG_DIR="$(pwd)"
+ELC="${ELC:-${LANG_DIR}/dist/platform/elc}"
+RT="${LANG_DIR}/el-compiler/runtime"
+
+PROG="${1:?usage: build.sh <program.el> <out-binary>}"
+OUT="${2:?usage: build.sh <program.el> <out-binary>}"
+
+# swarm module load order (each may depend on those before it):
+#   worktrack   — durable work-tracking journal (no swarm deps)
+#   containment — the three containment rules (no swarm deps)
+#   primitives  — think/act/attend/intend/learn seam (no swarm deps)
+#   ccr         — per-worker compiled bounded context (depends: primitives)
+#   swarm       — orchestrator: fan-out/converge (depends: all above + thread)
+SWARM_MODULES="
+  swarm/worktrack.el
+  swarm/containment.el
+  swarm/primitives.el
+  swarm/reshape_surface.el
+  swarm/primitive_binding.el
+  swarm/primitive_seam.el
+  swarm/ccr.el
+  swarm/swarm.el
+"
+
+TMP_C="$(mktemp -t swarm_build.XXXXXX).c"
+COMBINED="$(mktemp -t swarm_combined.XXXXXX).el"
+
+cat runtime/thread.el runtime/channel.el $SWARM_MODULES "$PROG" > "$COMBINED"
+
+if ! "$ELC" "$COMBINED" > "$TMP_C" 2>/tmp/swarm.elc.err; then
+    echo "elc FAILED:" >&2
+    sed 's/^/    /' /tmp/swarm.elc.err >&2
+    rm -f "$TMP_C" "$COMBINED"
+    exit 1
+fi
+
+if ! cc -O2 -DHAVE_CURL -I "$RT" "$TMP_C" "$RT/el_runtime.c" -lcurl -lpthread -lm -o "$OUT" 2>/tmp/swarm.cc.err; then
+    echo "cc FAILED:" >&2
+    sed 's/^/    /' /tmp/swarm.cc.err >&2
+    rm -f "$TMP_C" "$COMBINED"
+    exit 1
+fi
+
+rm -f "$TMP_C" "$COMBINED"
+echo "built: $OUT"

lang/swarm/ccr.el

@@ -0,0 +1,153 @@
+// ccr.el — Compiled Context Routing for work distribution.
+//
+// The same spine as the API's vantage-read, applied per worker. Instead of
+// handing every worker the coordinator's full memory, CCR compiles a MINIMAL,
+// BOUNDED context package scoped to exactly one worker's input (CCR §5, "Compiled
+// Context Injection"; Swarm §9.3, "The Compiled Context Boundary as Security
+// Boundary").
+//
+// The pipeline is CCR §5.1: Retrieval -> Scoping -> Compilation -> (Injection,
+// which here is placing the package into the worker's task envelope).
+//
+//   1. Retrieval  — resolve the blueprint's knowledge refs + the input's salient
+//                   terms against the mind (primitive_attend).
+//   2. Scoping    — keep only what THIS input needs; drop everything else. A
+//                   worker never receives sibling inputs or unrelated memory.
+//   3. Compilation— compact to a CTX string within a token budget (lossless of
+//                   meaning, smaller in tokens): collapse blank runs, dedupe
+//                   lines, then bound to the budget.
+//
+// The package a worker receives is therefore (a) sufficient for its task and
+// (b) incapable of leaking what it was never given — the containment boundary
+// and the security boundary are the same object.
+
+// ── token budget helpers ─────────────────────────────────────────────────────
+
+// ccr_est_tokens — cheap token estimate (~4 chars/token).
+fn ccr_est_tokens(s: String) -> Int {
+    return str_len(s) / 4
+}
+
+// ccr_default_budget — default per-worker context budget in tokens.
+//   Override with CCR_TOKEN_BUDGET.
+fn ccr_default_budget() -> Int {
+    let b: String = env("CCR_TOKEN_BUDGET")
+    if str_eq(b, "") {
+        return 1200
+    }
+    return str_to_int(b)
+}
+
+// ── stage 3: compaction ──────────────────────────────────────────────────────
+
+// ccr_compact — collapse blank-line runs and drop exact duplicate lines, then
+// bound the result to `budget` tokens (truncate on a line boundary). Meaning is
+// preserved; token count falls (CCR §5.2).
+fn ccr_compact(text: String, budget: Int) -> String {
+    let lines: [String] = str_split_lines(text)
+    let n: Int = el_list_len(lines)
+    let seen: String = "\n"
+    let out: String = ""
+    let out_tokens = 0
+    let i = 0
+    while i < n {
+        let ln: String = str_trim(el_list_get(lines, i))
+        if str_eq(ln, "") {
+            let i = i + 1
+        } else {
+            let marker: String = "\n" + ln + "\n"
+            if str_contains(seen, marker) {
+                // duplicate line — skip
+                let i = i + 1
+            } else {
+                let seen = seen + ln + "\n"
+                let line_tokens: Int = ccr_est_tokens(ln) + 1
+                if out_tokens + line_tokens > budget {
+                    // budget exhausted — stop (bounded)
+                    let i = n
+                } else {
+                    let out = out + ln + "\n"
+                    let out_tokens = out_tokens + line_tokens
+                    let i = i + 1
+                }
+            }
+        }
+    }
+    return out
+}
+
+// ── stages 1+2: retrieve + scope ─────────────────────────────────────────────
+
+// ccr_retrieve_scoped — pull context relevant to this input and its blueprint
+// knowledge refs, scoped to a fraction of the budget so no single source floods
+// the package. Returns compacted retrieved text (may be empty if the mind is
+// unreachable — the input alone is still a valid minimal context).
+fn ccr_retrieve_scoped(blueprint: String, knowledge_refs: String, input_item: String, budget: Int) -> String {
+    let acc: String = ""
+    // knowledge_refs is a JSON array of query strings.
+    let m: Int = json_array_len(knowledge_refs)
+    let i = 0
+    while i < m {
+        let ref: String = json_array_get_string(knowledge_refs, i)
+        let hit: String = primitive_attend(ref, 3)
+        let acc = acc + "# ref:" + ref + "\n" + hit + "\n"
+        let i = i + 1
+    }
+    // the input's own salient text also seeds retrieval
+    let hit2: String = primitive_attend(input_item, 3)
+    let acc = acc + "# input-context\n" + hit2 + "\n"
+    // scope retrieval to ~60% of budget; the input itself gets the rest
+    let retr_budget: Int = (budget * 6) / 10
+    return ccr_compact(acc, retr_budget)
+}
+
+// ── ccr_compile — assemble the bounded per-worker context package ─────────────
+//
+//   blueprint      : task blueprint name
+//   knowledge_refs : JSON array of retrieval queries from the blueprint
+//   input_item     : THIS worker's single input (and nothing else)
+//   corr_id        : swarm correlation ID
+//   worker_id      : this worker's ID
+//   scope_token    : the worker's containment token (closed boundary)
+//
+// Returns a JSON package: { blueprint, corr_id, worker_id, scope_token,
+// input, knowledge, budget_tokens, compiled_tokens }. `knowledge` is compiled
+// and bounded; the package as a whole is bounded by budget.
+fn ccr_compile(blueprint: String, knowledge_refs: String, input_item: String,
+               corr_id: String, worker_id: String, scope_token: String) -> String {
+    let budget: Int = ccr_default_budget()
+    let knowledge: String = ccr_retrieve_scoped(blueprint, knowledge_refs, input_item, budget)
+
+    let kv: [String] = el_list_empty()
+    let kv = el_list_append(kv, "blueprint")
+    let kv = el_list_append(kv, blueprint)
+    let kv = el_list_append(kv, "corr_id")
+    let kv = el_list_append(kv, corr_id)
+    let kv = el_list_append(kv, "worker_id")
+    let kv = el_list_append(kv, worker_id)
+    let kv = el_list_append(kv, "input")
+    let kv = el_list_append(kv, input_item)
+    let kv = el_list_append(kv, "knowledge")
+    let kv = el_list_append(kv, knowledge)
+    let kv = el_list_append(kv, "budget_tokens")
+    let kv = el_list_append(kv, int_to_str(budget))
+    let pkg: String = json_build_object(kv)
+    // stamp the scope token as a nested object, and the measured size
+    let pkg2: String = json_set(pkg, "scope_token", scope_token)
+    let compiled_tokens: Int = ccr_est_tokens(pkg2)
+    let pkg3: String = json_set(pkg2, "compiled_tokens", int_to_str(compiled_tokens))
+    return pkg3
+}
+
+// ccr_within_budget — did the compiled package stay within its budget?
+// (Retrieval is bounded to 60% and the input is small; this asserts the whole
+// package is bounded — the property distribution relies on.)
+fn ccr_within_budget(pkg: String) -> Bool {
+    let budget: Int = str_to_int(json_get_string(pkg, "budget_tokens"))
+    let compiled: Int = str_to_int(json_get_string(pkg, "compiled_tokens"))
+    // allow a small envelope for JSON framing overhead
+    if compiled <= budget + 200 {
+        return true
+    }
+    return false
+}

lang/swarm/containment.el

@@ -0,0 +1,181 @@
+// containment.el — the Swarm Architecture containment rules, enforced.
+//
+// "These rules are not conventions. They are enforced by the runtime."
+// (Swarm Architecture §3.2). The three rules that make bounded parallelism —
+// and therefore location-independent distribution — safe:
+//
+//   Rule 1: a worker may NOT join another swarm.
+//   Rule 2: a worker may NOT initiate a new swarm.
+//   Rule 3: a worker may NOT communicate laterally with sibling workers.
+//
+// Enforcement is by SCOPE TOKEN. When a swarm fans out, the coordinator mints a
+// swarm scope token and stamps a distinct worker scope token into each worker's
+// task envelope. Any attempt to create or join a swarm checks the caller's
+// token: if the caller already holds a WORKER token, the operation is rejected.
+// Rule 3 is enforced structurally elsewhere — workers share no mutable state and
+// the only channels they hold are the vertical result path — but this module
+// provides the explicit lateral-edge check for the execution tree.
+//
+// A scope token is a JSON object: {"kind":"coordinator|worker","swarm":"<corr>",
+// "worker":"<id-or-empty>","depth":"<n>"}.
+
+// ── Token minting ────────────────────────────────────────────────────────────
+
+// CAPABILITIES. A scope token carries a `caps` set — the authority it holds.
+// This is an AUTHORITY gate, not a health gate: capability is decided at mint
+// time and cannot be acquired at runtime. Engram-WRITE (op_write/op_relate/
+// op_supersede -> POST /api/nodes, /api/edges, DELETE) and dharma_emit are
+// @manager-ONLY capabilities — exactly the VBD rule that only the orchestrator
+// mutates global state. The orchestrator's token carries them; a worker's token
+// NEVER does. A worker is therefore STRUCTURALLY UNABLE to mutate global engram
+// state, regardless of engram health.
+fn cap_orchestrator() -> String { return "engram:read,engram:write,dharma:emit,state:write" }
+fn cap_worker() -> String { return "engram:read" }
+
+// containment_coordinator_token — the token the orchestrator (@manager) holds.
+// Depth 0. Carries the engram-WRITE + dharma-emit capabilities (@manager-only).
+fn containment_coordinator_token(corr_id: String) -> String {
+    let kv: [String] = el_list_empty()
+    let kv = el_list_append(kv, "kind")
+    let kv = el_list_append(kv, "coordinator")
+    let kv = el_list_append(kv, "swarm")
+    let kv = el_list_append(kv, corr_id)
+    let kv = el_list_append(kv, "worker")
+    let kv = el_list_append(kv, "")
+    let kv = el_list_append(kv, "depth")
+    let kv = el_list_append(kv, "0")
+    let kv = el_list_append(kv, "caps")
+    let kv = el_list_append(kv, cap_orchestrator())
+    return json_build_object(kv)
+}
+
+// containment_worker_token — the token stamped into a worker's envelope. Depth 1.
+// A closed boundary: forbids opening/joining swarms AND carries ONLY the
+// engram:READ capability — no engram:write, no dharma:emit. Read-only against the
+// full engram; may write only its own local geometry (its returned result).
+fn containment_worker_token(corr_id: String, worker_id: String) -> String {
+    let kv: [String] = el_list_empty()
+    let kv = el_list_append(kv, "kind")
+    let kv = el_list_append(kv, "worker")
+    let kv = el_list_append(kv, "swarm")
+    let kv = el_list_append(kv, corr_id)
+    let kv = el_list_append(kv, "worker")
+    let kv = el_list_append(kv, worker_id)
+    let kv = el_list_append(kv, "depth")
+    let kv = el_list_append(kv, "1")
+    let kv = el_list_append(kv, "caps")
+    let kv = el_list_append(kv, cap_worker())
+    return json_build_object(kv)
+}
+
+// containment_has_cap — does this token carry capability `cap`?
+fn containment_has_cap(token: String, cap: String) -> Bool {
+    return str_contains(json_get_string(token, "caps"), cap)
+}
+
+// ── Rule checks (return "" on allow, or a rejection reason string) ───────────
+
+// containment_check_open — may the holder of `token` OPEN a new swarm?
+// Enforces Rule 2 (a worker may not initiate a new swarm). Only a coordinator
+// token, or an absent token (top-level process), may open one.
+fn containment_check_open(token: String) -> String {
+    if str_eq(token, "") {
+        return ""
+    }
+    let kind: String = json_get_string(token, "kind")
+    if str_eq(kind, "worker") {
+        return "CONTAINMENT rule 2: a swarm worker may not initiate a new swarm (worker=" + json_get_string(token, "worker") + " swarm=" + json_get_string(token, "swarm") + ")"
+    }
+    return ""
+}
+
+// containment_check_join — may the holder of `token` JOIN swarm `target_corr`?
+// Enforces Rule 1 (a worker may not join another swarm). A worker already bound
+// to swarm A may not register into swarm B; and a worker may not re-join at all.
+fn containment_check_join(token: String, target_corr: String) -> String {
+    if str_eq(token, "") {
+        return ""
+    }
+    let kind: String = json_get_string(token, "kind")
+    if str_eq(kind, "worker") {
+        return "CONTAINMENT rule 1: a swarm worker may not join another swarm (worker=" + json_get_string(token, "worker") + " bound-swarm=" + json_get_string(token, "swarm") + " attempted-swarm=" + target_corr + ")"
+    }
+    return ""
+}
+
+// containment_check_lateral — may `from_token` open a communication edge to a
+// sibling worker `to_worker_id`? Enforces Rule 3 (no lateral communication).
+// The only permitted edges are vertical: worker->coordinator and
+// coordinator->worker. Any worker->worker edge is rejected.
+fn containment_check_lateral(from_token: String, to_worker_id: String) -> String {
+    let kind: String = json_get_string(from_token, "kind")
+    if str_eq(kind, "worker") {
+        if str_eq(to_worker_id, "") {
+            // empty target = the coordinator (vertical) — allowed
+            return ""
+        }
+        return "CONTAINMENT rule 3: a swarm worker may not communicate laterally with sibling workers (from=" + json_get_string(from_token, "worker") + " to=" + to_worker_id + ")"
+    }
+    return ""
+}
+
+// containment_check_engram_write — RULE 4: only a token carrying the
+// engram:write capability (the orchestrator's) may mutate global engram state.
+// A worker token (engram:read only) is REJECTED — the authority gate. Reuses the
+// exact scope-token mechanism as Rule 2's open-denial. Returns "" on allow, or a
+// rejection reason. This is an AUTHORITY gate: it does not consult engram health.
+fn containment_check_engram_write(token: String, op: String) -> String {
+    if containment_has_cap(token, "engram:write") {
+        return ""
+    }
+    return "CONTAINMENT rule 4: engram-write is @manager-only — a worker is read-only against the engram and may not mutate global state (op=" + op + " kind=" + json_get_string(token, "kind") + " worker=" + json_get_string(token, "worker") + " caps=" + json_get_string(token, "caps") + ")"
+}
+
+// containment_check_dharma_emit — the same @manager-only rule for dharma_emit,
+// grounding Rule 4 in VBD: global-state mutations (engram-write, dharma-emit) are
+// orchestrator-only, checked by the one capability mechanism.
+fn containment_check_dharma_emit(token: String) -> String {
+    if containment_has_cap(token, "dharma:emit") {
+        return ""
+    }
+    return "CONTAINMENT rule 4: dharma_emit is @manager-only (kind=" + json_get_string(token, "kind") + ")"
+}
+
+// ── Enforcement helpers ──────────────────────────────────────────────────────
+
+// containment_allows_open — Bool convenience over containment_check_open.
+fn containment_allows_open(token: String) -> Bool {
+    return str_eq(containment_check_open(token), "")
+}
+
+// containment_is_worker — is this a worker-scoped (closed-boundary) token?
+fn containment_is_worker(token: String) -> Bool {
+    return str_eq(json_get_string(token, "kind"), "worker")
+}
+
+// containment_guard_open — assert a swarm may be opened under this token.
+// Returns "" if allowed, or records a CONTAINMENT violation to the work-tracking
+// journal and returns the reason. Callers must abort on a non-empty return.
+fn containment_guard_open(token: String, corr_id: String) -> String {
+    let reason: String = containment_check_open(token)
+    if str_eq(reason, "") {
+        return ""
+    }
+    let p: String = json_set_str("{}", "reason", reason)
+    worktrack_append("containment.violation", corr_id, "open", p)
+    return reason
+}
+
+// containment_guard_engram_write — assert a token may mutate global engram state
+// (Rule 4). Returns "" if allowed; otherwise journals a containment.violation and
+// returns the reason. The write path MUST abort on a non-empty return.
+fn containment_guard_engram_write(token: String, corr_id: String, op: String) -> String {
+    let reason: String = containment_check_engram_write(token, op)
+    if str_eq(reason, "") {
+        return ""
+    }
+    let p0: String = json_set_str("{}", "reason", reason)
+    let p1: String = json_set_str(p0, "op", op)
+    worktrack_append("containment.violation", corr_id, "engram-write", p1)
+    return reason
+}

lang/swarm/primitive_binding.el

@@ -0,0 +1,49 @@
+// primitive_binding.el — THE ONE FLIP POINT.
+//
+// This file is the single seam between the swarm and the real agentic
+// primitives. Binding the reshape's decorated primitives is a one-line change
+// HERE and nothing else changes anywhere in the swarm.
+//
+// The api-reshape agent (wt/api-reshape) is wiring the primitives as DECORATED
+// El on the dharma_* event bus over the engram — think/attend/learn/ground/assert
+// become decorated fns that emit afferent events onto the bus. The moment they
+// land, flip `bound_think` (and its siblings) to call them.
+//
+//   TODAY  (stub fallback, compiles + runs now against :8901):
+//       fn bound_think(...) { return primitive_think(ctx, instruction) }
+//
+//   THE FLIP (when reshape's decorated primitives land — one line each):
+//       fn bound_think(...) { return think(ctx, instruction) }   // decorated, on dharma bus
+//
+// Keep the stub as fallback: `bound_think` is only reached when the seam mode is
+// "decorated" (SWARM_PRIMITIVE_SEAM=decorated). Until you flip these bodies AND
+// set that env, the harness runs entirely on the hermetic stub.
+
+// bound_think — BOUND to the reshape's proven decorated `think` (op_think),
+// real cognition over the engram geometry. The worker's CCR slice carries a
+// NODE-ID anchor in ctx.input (free-text anchors return "geometry unavailable");
+// think re-origins at that node's region under the faculty and returns a real
+// 768-dim gradient.
+fn bound_think(ctx: String, instruction: String) -> String {
+    let anchor: String = json_get_string(ctx, "input")
+    let faculty: String = json_get_string(ctx, "faculty")
+    return op_think(anchor, faculty)
+}
+
+// bound_attend — BOUND to the reshape's op_attend (POST /api/attend). Needs the
+// gate-1 write-healthy clone; falls back to the read-side attend otherwise.
+fn bound_attend(query: String, limit: Int) -> String {
+    if str_eq(env("SWARM_WRITE_HEALTHY"), "1") {
+        return op_attend(query, "self")
+    }
+    return primitive_attend(query, limit)
+}
+
+// bound_learn — BOUND to the reshape's op_learn (correspondence-beat). Needs the
+// gate-1 write-healthy clone; falls back to the opt-in journal-only learn.
+fn bound_learn(corr_id: String, observation: String) -> String {
+    if str_eq(env("SWARM_WRITE_HEALTHY"), "1") {
+        return op_learn(observation, "induce")
+    }
+    return primitive_learn(corr_id, observation)
+}

lang/swarm/primitive_seam.el

@@ -0,0 +1,55 @@
+// primitive_seam.el — the configurable primitive seam + telemetry.
+//
+// One switch selects where a worker's primitive invocation goes:
+//   SWARM_PRIMITIVE_SEAM=stub       (default) — hermetic in-process think.
+//   SWARM_PRIMITIVE_SEAM=decorated            — the reshape's decorated
+//                                               primitives on the dharma bus
+//                                               (see primitive_binding.el).
+//
+// Every seam invocation is an AFFERENT signal — a primitive call travelling
+// toward the manager. The seam stamps telemetry onto each thought (seam_mode +
+// one afferent tick) so the coordinator can aggregate afferent counters across
+// the swarm without any shared mutable state (containment-safe: counts ride the
+// vertical result path, not a shared bus register).
+
+// seam_mode — "stub" (default) or "decorated".
+fn seam_mode() -> String {
+    let m: String = env("SWARM_PRIMITIVE_SEAM")
+    if str_eq(m, "decorated") {
+        return "decorated"
+    }
+    return "stub"
+}
+
+// seam_think — route a worker's `think` through the configured seam and stamp
+// telemetry. Returns the thought JSON augmented with:
+//   seam_mode : which side of the seam served this call
+//   afferent  : "1" — one afferent primitive signal was emitted
+fn seam_think(ctx: String, instruction: String) -> String {
+    let mode: String = seam_mode()
+    let thought: String = ""
+    if str_eq(mode, "decorated") {
+        let thought = bound_think(ctx, instruction)
+    } else {
+        let thought = primitive_think(ctx, instruction)
+    }
+    let t1: String = json_set_str(thought, "seam_mode", mode)
+    let t2: String = json_set_str(t1, "afferent", "1")
+    return t2
+}
+
+// seam_attend / seam_learn — same seam for the other primitives (used when a
+// blueprint retrieves or writes through the bus).
+fn seam_attend(query: String, limit: Int) -> String {
+    if str_eq(seam_mode(), "decorated") {
+        return bound_attend(query, limit)
+    }
+    return primitive_attend(query, limit)
+}
+
+fn seam_learn(corr_id: String, observation: String) -> String {
+    if str_eq(seam_mode(), "decorated") {
+        return bound_learn(corr_id, observation)
+    }
+    return primitive_learn(corr_id, observation)
+}

lang/swarm/primitives.el

@@ -0,0 +1,94 @@
+// primitives.el — the agentic primitive SEAM the swarm composes over.
+//
+// The swarm is orchestration OVER the five CCR primitives, not a replacement for
+// them (CCR §2, "The Five Primitives / The Execution Cycle"): a worker executes
+// its task blueprint as attend -> think -> intend -> act -> learn against its
+// compiled, bounded context.
+//
+// This file is the SEAM. The parallel API-surface reshape exposes the canonical
+// primitive tools; when it lands, bind each primitive below to the reshaped
+// implementation (see PRIMITIVE_BINDING). Until then these are thin, engram-
+// backed fallbacks so the swarm — its fan-out, containment, CCR context
+// compilation, convergence, and work-tracking — is fully exercisable today.
+//
+// Contract: every primitive takes and returns String (JSON where structured), so
+// any primitive is directly threadable via thread.el's spawn (which runs
+// top-level (String)->String El fns).
+//
+// PRIMITIVE_BINDING: to bind the reshape's real tools, replace each fallback body
+// with a call to the reshaped El fn / API endpoint. Signatures here are the
+// stable contract the swarm depends on; keep them.
+
+// ── attend — retrieve the minimal relevant context for a focus ───────────────
+// Vantage-read: pull only what this focus needs from the mind. Backed by the
+// engram's spreading-activation retrieval.
+fn primitive_attend(query: String, limit: Int) -> String {
+    if str_eq(query, "") {
+        return "[]"
+    }
+    // Location-independent worker model: when an engram daemon is configured,
+    // retrieve over HTTP (the worker may run anywhere). POST /api/search
+    // {query,limit,_auth}. Falls back to the in-process store otherwise.
+    let url: String = env("ENGRAM_URL")
+    if str_eq(url, "") {
+        return engram_activate(query, limit)
+    }
+    let kv: [String] = el_list_empty()
+    let kv = el_list_append(kv, "query")
+    let kv = el_list_append(kv, query)
+    let body0: String = json_build_object(kv)
+    let body1: String = json_set(body0, "limit", int_to_str(limit))
+    let body2: String = json_set_str(body1, "_auth", env("ENGRAM_API_KEY"))
+    return http_post(url + "/api/search", body2)
+}
+
+// ── think — reason over the compiled context ─────────────────────────────────
+// In production this routes to a model (CCR dynamic model selection). Here it is
+// a deterministic, hermetic transform so swarm behaviour is testable without an
+// external model: it echoes a structured verdict derived from the context. The
+// binding point for a real model is explicit.
+fn primitive_think(compiled_ctx: String, instruction: String) -> String {
+    // PRIMITIVE_BINDING: replace with the reshape's think() (model inference).
+    let kv: [String] = el_list_empty()
+    let kv = el_list_append(kv, "instruction")
+    let kv = el_list_append(kv, instruction)
+    let kv = el_list_append(kv, "ctx_bytes")
+    let kv = el_list_append(kv, int_to_str(str_len(compiled_ctx)))
+    let kv = el_list_append(kv, "conclusion")
+    let kv = el_list_append(kv, "reasoned:" + instruction)
+    return json_build_object(kv)
+}
+
+// ── intend — form a bounded plan/decision from a thought ─────────────────────
+fn primitive_intend(thought: String) -> String {
+    let concl: String = json_get_string(thought, "conclusion")
+    let kv: [String] = el_list_empty()
+    let kv = el_list_append(kv, "intent")
+    let kv = el_list_append(kv, concl)
+    return json_build_object(kv)
+}
+
+// ── act — execute a bounded effect and return its result ─────────────────────
+// Workers defer real side-effects to the coordinator (idempotency requirement,
+// Swarm §7.3). Here act produces an artifact-shaped result the coordinator
+// collects during convergence.
+fn primitive_act(intent: String, input_item: String) -> String {
+    let kv: [String] = el_list_empty()
+    let kv = el_list_append(kv, "acted_on")
+    let kv = el_list_append(kv, input_item)
+    let kv = el_list_append(kv, "via")
+    let kv = el_list_append(kv, json_get_string(intent, "intent"))
+    return json_build_object(kv)
+}
+
+// ── learn — record an observation into the mind, tagged by correlation ID ────
+// Append-only, naturally idempotent (Swarm §7.3). Best-effort: a worker that
+// cannot reach the mind still returns its result.
+fn primitive_learn(corr_id: String, observation: String) -> String {
+    let url: String = env("ENGRAM_URL")
+    if str_eq(url, "") {
+        return ""
+    }
+    let content: String = "swarm-worker-obs corr=" + corr_id + " :: " + observation
+    return engram_node(content, "Memory", 0.4)
+}

lang/swarm/reshape_surface.el

@@ -0,0 +1,103 @@
+// reshape_surface.el — the api-reshape agent's PROVEN decorated primitives,
+// composed into the swarm build to bind real cognition.
+//
+// PROVENANCE: these fns are the reshape's surface at wt/api-reshape @ d4f401d
+// ("reshape: decorator-as-seam — port @route codegen, prove decorate->serve,
+// rewrite surface as decorated El"), verified live against
+// engram.cognition-20260814. Copied verbatim (read/cognition ops only) so the
+// swarm binds the REAL primitives, not a reimplementation. The write ops
+// (op_write/op_relate/op_supersede/op_ground) are intentionally NOT composed
+// here — they exercise the persist_node write path that needs the gate-1
+// write-healthy clone; the swarm's proven run is read-cognition (think/read).
+//
+// Ops route to the ENGRAM over ENGRAM_URL — pinned by THIS worktree's .nsbx-env
+// to the :8901 swarm clone (never the reshape agent's :8900). Separate clones,
+// no collision.
+
+fn engram_url() -> String {
+    let u: String = env("ENGRAM_URL")
+    if str_eq(u, "") { return "http://127.0.0.1:8900" }
+    return u
+}
+fn engram_key() -> String {
+    let k: String = env("ENGRAM_API_KEY")
+    if str_eq(k, "") { return "sbx-dev-api-reshape" }
+    return k
+}
+fn SELF_KEY() -> String { return "kn-efeb4a5b-5aff-4759-8a97-7233099be6ee" }
+fn VALUES_KEY() -> String { return "kn-5b606390-a52d-4ca2-8e0e-eba141d13440" }
+
+// self/values name -> keystone id; anything else passes through unchanged.
+fn resolve_named(v: String) -> String {
+    if str_eq(v, "self") { return SELF_KEY() }
+    if str_eq(v, "neuron") { return SELF_KEY() }
+    if str_eq(v, "values") { return VALUES_KEY() }
+    if str_eq(v, "values_hub") { return VALUES_KEY() }
+    return v
+}
+
+// read — THE VANTAGE-READ. Re-origin at a point + aperture -> a BOUNDED slice.
+fn op_read(vantage: String, typ: String, k: Int) -> String {
+    let vid: String = resolve_named(vantage)
+    if str_eq(typ, "edges") {
+        return http_get(engram_url() + "/api/neighbors/" + vid)
+    }
+    if str_starts_with(vid, "kn-") {
+        return http_get(engram_url() + "/api/neighbors/" + vid)
+    }
+    return http_get(engram_url() + "/api/search?q=" + url_encode(vid) + "&limit=" + int_to_str(k))
+}
+
+// think — THE ONE OPERATION. anchor (node ids) steered by faculty -> gradient.
+fn op_think(seeds: String, faculty: String) -> String {
+    let s: String = resolve_named(seeds)
+    let f: String = if str_eq(faculty, "") { "reason" } else { faculty }
+    return http_get(engram_url() + "/api/think?seeds=" + url_encode(s) + "&faculty=" + f)
+}
+
+// attend — aim attention at a region. (POST — needs a write-healthy clone.)
+fn op_attend(node: String, observer: String) -> String {
+    let n: String = resolve_named(node)
+    let o: String = if str_eq(observer, "") { SELF_KEY() } else { resolve_named(observer) }
+    let body: String = "{\"_auth\":\"" + engram_key() + "\",\"node\":\"" + n
+        + "\",\"observer\":\"" + o + "\",\"salience\":\"0.6\"}"
+    return http_post_json(engram_url() + "/api/attend", body)
+}
+
+fn identity_typed(t: String) -> Bool {
+    if str_eq(t, "self") { return true }
+    if str_eq(t, "values") { return true }
+    return false
+}
+fn type_to_node_type(t: String) -> String {
+    if str_eq(t, "knowledge") { return "Knowledge" }
+    if str_eq(t, "artifact") { return "Artifact" }
+    if str_eq(t, "backlog") { return "WorkItem" }
+    if str_eq(t, "process") { return "Process" }
+    if str_eq(t, "state") { return "InternalStateEvent" }
+    return "Memory"
+}
+
+// write — add a node (POST /api/nodes). Identity types refused. This is a
+// global-engram MUTATION — @manager-only (Rule 4); never called on a worker path.
+// (Reshape's op_write, with json_escape -> the available json_escape_string.)
+fn op_write(content: String, typ: String, importance: Float) -> String {
+    if str_eq(content, "") { return "{\"error\":\"write: content required\"}" }
+    if identity_typed(typ) {
+        return "{\"error\":\"write type=" + typ + " is write-protected -> intentional-cultivation\"}"
+    }
+    let body: String = "{\"_auth\":\"" + engram_key() + "\",\"content\":\"" + json_escape_string(content)
+        + "\",\"node_type\":\"" + type_to_node_type(typ) + "\",\"tier\":\"Working\",\"importance\":"
+        + float_to_str(importance) + "}"
+    return http_post_json(engram_url() + "/api/nodes", body)
+}
+
+// learn — the reflexive correspondence-beat: calibrate the steering-prior.
+// (POST — needs a write-healthy clone.)
+fn op_learn(seeds: String, faculty: String) -> String {
+    let s: String = resolve_named(seeds)
+    let f: String = if str_eq(faculty, "") { "induce" } else { faculty }
+    let body: String = "{\"_auth\":\"" + engram_key() + "\",\"seeds\":\"" + s
+        + "\",\"faculty\":\"" + f + "\",\"keystone\":\"false\"}"
+    return http_post_json(engram_url() + "/api/correspondence-beat", body)
+}

lang/swarm/swarm.el

@@ -0,0 +1,483 @@
+// swarm.el — the swarm orchestrator: bounded parallel agent execution.
+//
+// Implements Swarm Architecture's single pattern — fan out, execute independently,
+// converge — on El's NATIVE concurrency (thread.el spawn/join). No external
+// orchestrator: a swarm is a coordinator (this file, the main thread) that mints
+// a correlation identity, compiles a bounded CCR context per worker, dispatches
+// workers as native pthreads, tracks every unit of work, and converges the
+// results before returning control to the parent step.
+//
+// The five properties of every swarm (Swarm §2.1) are all present:
+//   parent step        -> swarm_run is called from one process step
+//   task blueprint     -> `blueprint` name + knowledge refs, run by every worker
+//   input set          -> `inputs_json`, one item per worker
+//   convergence        -> `strategy` in config (collect|merge|vote|reduce)
+//   correlation ID     -> minted here, threaded through tracking + every worker
+//
+// Containment (Swarm §3) is enforced: the caller must hold a coordinator/absent
+// token to open a swarm (Rule 2), each worker is stamped a closed worker token
+// (Rules 1+3), and workers share no mutable state (the coordinator is the only
+// journal writer).
+
+// ── worker entry — the top-level (String)->String fn native threads run ──────
+//
+// Every El fn compiles to a global C symbol; spawn() resolves this by name via
+// dlsym and runs it in a pthread. The envelope carries everything the worker is
+// permitted to see — its compiled context and nothing else (§9.3).
+//
+// Returns a result JSON: {worker_id, status:"completed"|"failed", output|error}.
+fn swarm_worker_entry(envelope_json: String) -> String {
+    let worker_id: String = json_get_string(envelope_json, "worker_id")
+    let ctx: String = json_get_raw(envelope_json, "ctx")
+
+    // The worker holds a CLOSED worker token (Rules 1+3): it shares no state
+    // with siblings and may not open/join a swarm. That boundary is enforced at
+    // the point of attempt — swarm_run rejects any swarm opened under a worker
+    // token (Rule 2). A worker simply executing its blueprint is not opening a
+    // swarm, so it proceeds. Its only outward edge is this returned result
+    // (the vertical worker->coordinator path).
+    let out: String = swarm_run_blueprint(ctx)
+    // A worker reports failed iff its blueprint signalled failure. This is the
+    // vertical status edge the coordinator reads during convergence (§4.3, §7).
+    let bstatus: String = json_get_string(out, "blueprint_status")
+    let status: String = "completed"
+    if str_eq(bstatus, "failed") {
+        let status = "failed"
+    }
+    let kv: [String] = el_list_empty()
+    let kv = el_list_append(kv, "worker_id")
+    let kv = el_list_append(kv, worker_id)
+    let kv = el_list_append(kv, "status")
+    let kv = el_list_append(kv, status)
+    let res: String = json_build_object(kv)
+    return json_set(res, "output", out)
+}
+
+// swarm_run_blueprint — execute the task blueprint over a compiled context.
+// The default blueprint is the CCR execution cycle: think -> intend -> act over
+// the worker's bounded context. Specialise by dispatching on
+// json_get_string(ctx,"blueprint"). Idempotent: reads ctx, writes only its
+// returned output (§7.3).
+fn swarm_run_blueprint(ctx: String) -> String {
+    let blueprint: String = json_get_string(ctx, "blueprint")
+    let input_item: String = json_get_string(ctx, "input")
+    let knowledge: String = json_get_string(ctx, "knowledge")
+
+    // classify — deterministic verdict for the `vote` convergence strategy:
+    // verdict is "long" if the input has >4 chars, else "short".
+    if str_eq(blueprint, "classify") {
+        let verdict: String = "short"
+        if str_len(input_item) > 4 {
+            let verdict = "long"
+        }
+        let kv: [String] = el_list_empty()
+        let kv = el_list_append(kv, "verdict")
+        let kv = el_list_append(kv, verdict)
+        let kv = el_list_append(kv, "blueprint_status")
+        let kv = el_list_append(kv, "ok")
+        return json_build_object(kv)
+    }
+
+    // faildemo — a worker that fails on inputs beginning with "x" (exercises the
+    // failure threshold + partial convergence path). Idempotent, side-effect-free.
+    if str_eq(blueprint, "faildemo") {
+        let st: String = "ok"
+        if str_starts_with(input_item, "x") {
+            let st = "failed"
+        }
+        return json_set_str("{}", "blueprint_status", st)
+    }
+
+    // cognize — REAL-COGNITION blueprint. Routes think through the seam (bound to
+    // op_think in decorated mode) over the worker's NODE-ID anchor, then derives a
+    // vote verdict from the gradient's confidence. In stub mode there is no
+    // gradient, so the verdict falls back to a deterministic slice hash — the
+    // same blueprint runs green on either side of the seam.
+    if str_eq(blueprint, "cognize") {
+        let thought: String = seam_think(ctx, "reason over " + input_item)
+        // Derive the vote verdict from the REAL gradient's support count
+        // (json_get_int, since n_support is numeric). Different anchors have
+        // different support -> genuine, cognition-driven vote diversity. In stub
+        // mode there is no gradient (n_support -> 0) -> "uncertain".
+        let nsup: Int = json_get_int(thought, "n_support")
+        let verdict: String = "uncertain"
+        if nsup >= 10 {
+            let verdict = "confident"
+        }
+        let ck: [String] = el_list_empty()
+        let ck = el_list_append(ck, "verdict")
+        let ck = el_list_append(ck, verdict)
+        let ck = el_list_append(ck, "blueprint_status")
+        let ck = el_list_append(ck, "ok")
+        let cout0: String = json_build_object(ck)
+        let cout1: String = json_set_str(cout0, "n_support", int_to_str(nsup))
+        let cout2: String = json_set_str(cout1, "seam_mode", json_get_string(thought, "seam_mode"))
+        return json_set_str(cout2, "afferent", json_get_string(thought, "afferent"))
+    }
+
+    // default (analyze_item): the CCR execution cycle think -> intend -> act,
+    // with `think` routed through the CONFIGURABLE PRIMITIVE SEAM. Telemetry
+    // (seam_mode + afferent tick) rides the worker's returned output.
+    let instruction: String = "process input: " + input_item
+    let thought: String = seam_think(ctx, instruction)
+    let intent: String = primitive_intend(thought)
+    let effect: String = primitive_act(intent, input_item)
+    let e1: String = json_set_str(effect, "blueprint_status", "ok")
+    let e2: String = json_set_str(e1, "seam_mode", json_get_string(thought, "seam_mode"))
+    let e3: String = json_set_str(e2, "afferent", json_get_string(thought, "afferent"))
+    return e3
+}
+
+// ── native-thread fan-out, bounded by concurrency, order-preserving ──────────
+//
+// parallel_map (thread.el) spawns ALL threads at once. The swarm honours the
+// blueprint's `concurrency` cap (§5.1: a resource constraint, not a parallelism
+// constraint — all items are processed, at most N at a time) by dispatching in
+// waves of N native threads, joining each wave before the next. Results are
+// returned in input order.
+fn swarm_fanout(worker_fn: String, envelopes: [String], concurrency: Int) -> [String] {
+    let n: Int = el_list_len(envelopes)
+    let cap: Int = concurrency
+    if cap < 1 {
+        let cap = 1
+    }
+    let results: [String] = el_list_empty()
+    let base = 0
+    while base < n {
+        // spawn a wave of up to `cap` workers
+        let tids: [String] = el_list_empty()
+        let k = 0
+        while k < cap {
+            let idx: Int = base + k
+            if idx < n {
+                let env_item: String = el_list_get(envelopes, idx)
+                let tid: Int = spawn(worker_fn, env_item)
+                let tids = el_list_append(tids, int_to_str(tid))
+            }
+            let k = k + 1
+        }
+        // join the wave in order
+        let j = 0
+        let jn: Int = el_list_len(tids)
+        while j < jn {
+            let tid: Int = str_to_int(el_list_get(tids, j))
+            let r: String = join(tid)
+            let results = el_list_append(results, r)
+            let j = j + 1
+        }
+        let base = base + cap
+    }
+    return results
+}
+
+// ── convergence strategies (Swarm §4.2) ──────────────────────────────────────
+
+// swarm_converge_collect — ordered list, no transformation.
+fn swarm_converge_collect(results: [String]) -> String {
+    let n: Int = el_list_len(results)
+    let arr: String = "[]"
+    let i = 0
+    while i < n {
+        let arr = json_array_push(arr, el_list_get(results, i))
+        let i = i + 1
+    }
+    return arr
+}
+
+// swarm_converge_merge — combine worker outputs into a single joined string.
+fn swarm_converge_merge(results: [String]) -> String {
+    let n: Int = el_list_len(results)
+    let merged: String = ""
+    let i = 0
+    while i < n {
+        let out: String = json_get_raw(el_list_get(results, i), "output")
+        if i > 0 {
+            let merged = merged + " | "
+        }
+        let merged = merged + out
+        let i = i + 1
+    }
+    return json_set_str("{}", "merged", merged)
+}
+
+// swarm_converge_vote — tally a field across worker outputs, pick the majority.
+// Each worker output is expected to carry a "verdict" string field.
+fn swarm_converge_vote(results: [String]) -> String {
+    let n: Int = el_list_len(results)
+    // Collect verdicts (no mutable tally: json_set can't update an existing key
+    // and there is no el_list_set). Then count each verdict by rescanning.
+    let verdicts: [String] = el_list_empty()
+    let i = 0
+    while i < n {
+        let out: String = json_get_raw(el_list_get(results, i), "output")
+        let v: String = json_get_string(out, "verdict")
+        if str_eq(v, "") {
+            let i = i + 1
+        } else {
+            let verdicts = el_list_append(verdicts, v)
+            let i = i + 1
+        }
+    }
+    // pick the verdict with the highest count (first-past-the-post)
+    let vn: Int = el_list_len(verdicts)
+    let best: String = ""
+    let bestc = 0
+    let a = 0
+    while a < vn {
+        let cand: String = el_list_get(verdicts, a)
+        // count occurrences of cand
+        let c = 0
+        let b = 0
+        while b < vn {
+            if str_eq(el_list_get(verdicts, b), cand) {
+                let c = c + 1
+            }
+            let b = b + 1
+        }
+        if c > bestc {
+            let bestc = c
+            let best = cand
+        }
+        let a = a + 1
+    }
+    let kv: [String] = el_list_empty()
+    let kv = el_list_append(kv, "winner")
+    let kv = el_list_append(kv, best)
+    let kv = el_list_append(kv, "votes")
+    let kv = el_list_append(kv, int_to_str(bestc))
+    return json_build_object(kv)
+}
+
+// swarm_converge_reduce — fold outputs into an accumulator (count + concat).
+fn swarm_converge_reduce(results: [String]) -> String {
+    let n: Int = el_list_len(results)
+    let acc: String = ""
+    let i = 0
+    while i < n {
+        let out: String = json_get_raw(el_list_get(results, i), "output")
+        let acc = acc + out
+        let i = i + 1
+    }
+    let kv: [String] = el_list_empty()
+    let kv = el_list_append(kv, "count")
+    let kv = el_list_append(kv, int_to_str(n))
+    let kv = el_list_append(kv, "accumulated")
+    let kv = el_list_append(kv, acc)
+    return json_build_object(kv)
+}
+
+// ratio_to_permille — parse a decimal ratio string ("1.0", "0.8") into an
+// integer per-mille (1000, 800) so failure thresholds use exact integer math.
+// (El float division is unreliable in this runtime — int_to_float(n)/int_to_float(n)
+// does not equal 1.0 — so the swarm deliberately avoids floats.)
+fn ratio_to_permille(s: String) -> Int {
+    if str_eq(s, "") {
+        return 1000
+    }
+    let parts: [String] = str_split(s, ".")
+    let whole: Int = str_to_int(el_list_get(parts, 0))
+    let permille: Int = whole * 1000
+    if el_list_len(parts) > 1 {
+        let frac_raw: String = el_list_get(parts, 1)
+        let frac3: String = str_slice(str_pad_right(frac_raw, 3, "0"), 0, 3)
+        let permille = permille + str_to_int(frac3)
+    }
+    return permille
+}
+
+// swarm_converge — dispatch on strategy name.
+fn swarm_converge(strategy: String, results: [String]) -> String {
+    if str_eq(strategy, "merge") {
+        return swarm_converge_merge(results)
+    }
+    if str_eq(strategy, "vote") {
+        return swarm_converge_vote(results)
+    }
+    if str_eq(strategy, "reduce") {
+        return swarm_converge_reduce(results)
+    }
+    // default: collect
+    return swarm_converge_collect(results)
+}
+
+// ── the ONLY global-engram write path (Rule 4, @manager-only) ────────────────
+//
+// Every engram mutation flows through here and is gated by the caller's token
+// capability. Only the orchestrator's token carries engram:write, so a worker
+// (engram:read only) calling this is DENIED by capability before any HTTP is
+// issued — structurally unable to mutate global engram state, regardless of
+// engram health. This is the curated-merge write: the orchestrator committing
+// the geometry it approved. Workers never reach a successful branch here.
+fn swarm_engram_write(token: String, corr_id: String, content: String, typ: String, importance: Float) -> String {
+    let deny: String = containment_guard_engram_write(token, corr_id, "engram.write")
+    if str_eq(deny, "") {
+        // authorized (orchestrator) — perform the write
+        let res: String = op_write(content, typ, importance)
+        let new_id: String = json_get_string(res, "id")
+        let cp: String = json_set_str("{}", "node_id", new_id)
+        worktrack_append("swarm.committed", corr_id, "orchestrator", cp)
+        return res
+    }
+    // denied by capability — return the rejection, no engram mutation performed
+    return json_set_str("{}", "denied", deny)
+}
+
+// ── the coordinator: fan out -> track -> converge ────────────────────────────
+//
+//   blueprint        : task blueprint name run by every worker
+//   knowledge_refs   : JSON array of retrieval queries for CCR compilation
+//   inputs_json      : JSON array of input items (one per worker)
+//   config_json      : { concurrency, strategy, min_success_ratio,
+//                        failure_action, caller_token }
+//
+// Returns: { corr_id, status:"completed"|"aborted", merged, report }.
+fn swarm_run(blueprint: String, knowledge_refs: String, inputs_json: String, config_json: String) -> String {
+    let corr_id: String = "swarm-" + uuid_v4()
+    let caller_token: String = json_get_raw(config_json, "caller_token")
+    let concurrency: Int = str_to_int(json_get_string(config_json, "concurrency"))
+    if concurrency < 1 {
+        let concurrency = 4
+    }
+    let strategy: String = json_get_string(config_json, "strategy")
+
+    // ── Containment Rule 2: only a coordinator/absent token may open a swarm ──
+    let deny: String = containment_guard_open(caller_token, corr_id)
+    if str_eq(deny, "") {
+        // allowed — proceed
+        let n: Int = json_array_len(inputs_json)
+
+        // swarm.created
+        let cp: String = json_set_str("{}", "blueprint", blueprint)
+        let cp2: String = json_set(cp, "input_count", int_to_str(n))
+        worktrack_append("swarm.created", corr_id, corr_id, cp2)
+
+        // build per-worker envelopes: worker token + CCR-compiled bounded context
+        let envelopes: [String] = el_list_empty()
+        let i = 0
+        while i < n {
+            let worker_id: String = corr_id + "/worker-" + int_to_str(i)
+            let input_item: String = json_array_get_string(inputs_json, i)
+            let wtoken: String = containment_worker_token(corr_id, worker_id)
+            let ctx: String = ccr_compile(blueprint, knowledge_refs, input_item, corr_id, worker_id, wtoken)
+            // envelope: only this worker's compiled context + its closed token
+            let ekv: [String] = el_list_empty()
+            let ekv = el_list_append(ekv, "worker_id")
+            let ekv = el_list_append(ekv, worker_id)
+            let ekv = el_list_append(ekv, "corr_id")
+            let ekv = el_list_append(ekv, corr_id)
+            let env0: String = json_build_object(ekv)
+            let env1: String = json_set(env0, "scope_token", wtoken)
+            let env2: String = json_set(env1, "ctx", ctx)
+            let envelopes = el_list_append(envelopes, env2)
+
+            let sp: String = json_set_str("{}", "input", input_item)
+            worktrack_append("worker.started", corr_id, worker_id, sp)
+            let i = i + 1
+        }
+
+        // ── native-thread fan-out (bounded) ──
+        let results: [String] = swarm_fanout("swarm_worker_entry", envelopes, concurrency)
+
+        // record per-worker terminal status + aggregate AFFERENT telemetry.
+        // Afferent counters (primitive signals travelling toward the @manager)
+        // are summed from the vertical result path — no shared bus register,
+        // so the aggregation is containment-safe.
+        let succ = 0
+        let afferent = 0
+        let seam_mode_seen: String = "stub"
+        let rn: Int = el_list_len(results)
+        let r = 0
+        while r < rn {
+            let res: String = el_list_get(results, r)
+            let wid: String = json_get_string(res, "worker_id")
+            let st: String = json_get_string(res, "status")
+            let out: String = json_get_raw(res, "output")
+            let aff: Int = str_to_int(json_get_string(out, "afferent"))
+            let afferent = afferent + aff
+            let sm: String = json_get_string(out, "seam_mode")
+            if str_eq(sm, "") {
+                let seam_mode_seen = seam_mode_seen
+            } else {
+                let seam_mode_seen = sm
+            }
+            if str_eq(st, "completed") {
+                let succ = succ + 1
+                worktrack_append("worker.completed", corr_id, wid, json_set_str("{}", "status", "completed"))
+            } else {
+                worktrack_append("worker.failed", corr_id, wid, json_set_str("{}", "error", json_get_string(res, "error")))
+            }
+            let r = r + 1
+        }
+
+        // swarm.converging
+        let vg: String = json_set("{}", "success_count", int_to_str(succ))
+        worktrack_append("swarm.converging", corr_id, corr_id, vg)
+
+        // swarm.telemetry — afferent counters observed by the @manager.
+        let tkv: [String] = el_list_empty()
+        let tkv = el_list_append(tkv, "seam_mode")
+        let tkv = el_list_append(tkv, seam_mode_seen)
+        let telem0: String = json_build_object(tkv)
+        let telem1: String = json_set_str(telem0, "afferent_think", int_to_str(afferent))
+        let telemetry: String = json_set_str(telem1, "results_received", int_to_str(rn))
+        worktrack_append("swarm.telemetry", corr_id, corr_id, telemetry)
+
+        // ── failure threshold (Swarm §4.3), integer per-mille math ──
+        // require succ/n >= min_success_ratio  <=>  succ*1000 >= permille*n
+        let permille: Int = ratio_to_permille(json_get_string(config_json, "min_success_ratio"))
+        let status: String = "completed"
+        if succ * 1000 < permille * n {
+            let status = "aborted"
+        }
+
+        if str_eq(status, "aborted") {
+            let ap: String = json_set_str("{}", "reason", "success ratio below min_success_ratio")
+            worktrack_append("swarm.aborted", corr_id, corr_id, ap)
+            let rep: String = worktrack_swarm_report(corr_id)
+            let ok: [String] = el_list_empty()
+            let ok = el_list_append(ok, "corr_id")
+            let ok = el_list_append(ok, corr_id)
+            let ok = el_list_append(ok, "status")
+            let ok = el_list_append(ok, "aborted")
+            let out0: String = json_build_object(ok)
+            return json_set(out0, "report", rep)
+        }
+
+        // ── converge ──
+        let merged: String = swarm_converge(strategy, results)
+        let dp: String = json_set_str("{}", "strategy", strategy)
+        worktrack_append("swarm.completed", corr_id, corr_id, dp)
+
+        // ── curated merge = the ONLY engram write path (Rule 4) ──
+        // With "commit":"1", the ORCHESTRATOR (its token carries engram:write)
+        // commits the approved merged geometry back to the engram. This is the
+        // single writer. Workers returned geometry; only the orchestrator writes.
+        let commit_id: String = ""
+        if str_eq(json_get_string(config_json, "commit"), "1") {
+            let orch_token: String = containment_coordinator_token(corr_id)
+            let cres: String = swarm_engram_write(orch_token, corr_id, "swarm-merge " + corr_id + " :: " + merged, "memory", 0.5)
+            let commit_id = json_get_string(cres, "id")
+        }
+
+        let rep2: String = worktrack_swarm_report(corr_id)
+        let ok2: [String] = el_list_empty()
+        let ok2 = el_list_append(ok2, "corr_id")
+        let ok2 = el_list_append(ok2, corr_id)
+        let ok2 = el_list_append(ok2, "status")
+        let ok2 = el_list_append(ok2, "completed")
+        let out1: String = json_build_object(ok2)
+        let out2: String = json_set(out1, "report", rep2)
+        let out3: String = json_set(out2, "merged", merged)
+        let out4: String = json_set(out3, "telemetry", telemetry)
+        return json_set_str(out4, "committed_node", commit_id)
+    }
+    // ── denied: caller was a worker trying to open a swarm (Rule 2) ──
+    let dkv: [String] = el_list_empty()
+    let dkv = el_list_append(dkv, "corr_id")
+    let dkv = el_list_append(dkv, corr_id)
+    let dkv = el_list_append(dkv, "status")
+    let dkv = el_list_append(dkv, "denied")
+    let dkv = el_list_append(dkv, "error")
+    let dkv = el_list_append(dkv, deny)
+    return json_build_object(dkv)
+}

lang/swarm/tests/harness_local_swarm.el

@@ -0,0 +1,88 @@
+// harness_local_swarm.el — LOCAL-SWARM INTEGRATION HARNESS.
+//
+// Proves the FULL local-swarm mechanics end-to-end, TODAY, on the isolated
+// engram clone (:8901), with the primitive seam pointed at the hermetic stub.
+// The moment the api-reshape agent lands the decorated primitives on the
+// dharma bus, binding is ONE flip (primitive_binding.el) + SWARM_PRIMITIVE_SEAM=
+// decorated — this same harness then runs the bound path with no other change.
+//
+// The @manager (the coordinator) fans out N native El worker threads at real
+// concurrency, each given a CCR-scoped engram slice, each invoking the primitive
+// seam (think over its slice), enforces all three containment rules, converges
+// (vote AND reduce), work-tracks durably, and observes afferent telemetry.
+//
+// Run with the sandbox env sourced (ENGRAM_URL=:8901) to also exercise CCR
+// retrieval against the real (isolated) mind; runs fully without it too.
+
+fn ok(label: String, cond: Bool, fails: Int) -> Int {
+    if cond { print("  ok    " + label); return fails }
+    print("  FAIL  " + label); return fails + 1
+}
+
+fn main() -> Int {
+    let fails = 0
+    print("== LOCAL-SWARM INTEGRATION HARNESS (seam=" + seam_mode() + ") ==")
+
+    // 8 independent slices, real concurrency of 4 (2 waves of native pthreads).
+    let inputs: String = "[\"billing\",\"payments\",\"ledger\",\"invoicing\",\"tax\",\"payroll\",\"audit\",\"fx\"]"
+    let refs: String = "[\"Volatility-Based Decomposition\"]"
+
+    // ── A) fan-out / converge at real concurrency (reduce) ──
+    let cfg_r: String = "{\"concurrency\":\"4\",\"strategy\":\"reduce\",\"min_success_ratio\":\"1.0\"}"
+    let rr: String = swarm_run("analyze_item", refs, inputs, cfg_r)
+    let fails = ok("swarm completed at concurrency=4 over 8 native-thread workers", str_eq(json_get_string(rr, "status"), "completed"), fails)
+    let corr: String = json_get_string(rr, "corr_id")
+    let merged_r: String = json_get_raw(rr, "merged")
+    let fails = ok("reduce converged all 8 worker outputs", str_to_int(json_get_string(merged_r, "count")) == 8, fails)
+
+    // ── B) afferent telemetry observed by the @manager ──
+    let telem: String = json_get_raw(rr, "telemetry")
+    let aff: Int = str_to_int(json_get_string(telem, "afferent_think"))
+    let seen_mode: String = json_get_string(telem, "seam_mode")
+    let fails = ok("afferent think-signals counted = 8 (one per worker)", aff == 8, fails)
+    let fails = ok("telemetry records the active seam mode", str_eq(seen_mode, seam_mode()), fails)
+    let telem_recs: Int = worktrack_count_kind(corr, "swarm.telemetry")
+    let fails = ok("telemetry durably journalled", telem_recs == 1, fails)
+
+    // ── C) CCR scoping + non-leak per worker ──
+    let wt: String = containment_worker_token(corr, corr + "/worker-3")
+    let ctx3: String = ccr_compile("analyze_item", refs, "invoicing", corr, corr + "/worker-3", wt)
+    let fails = ok("CCR context bounded within token budget", ccr_within_budget(ctx3), fails)
+    let fails = ok("CCR context carries THIS slice", str_eq(json_get_string(ctx3, "input"), "invoicing"), fails)
+    let leaks: Bool = str_contains(ctx3, "payroll") || str_contains(ctx3, "audit")
+    let fails = ok("CCR context does NOT leak sibling slices (security boundary)", !leaks, fails)
+
+    // ── D) all three containment rules ──
+    let deny: String = containment_check_open(wt)
+    let fails = ok("Rule 2: worker token may not OPEN a swarm", !str_eq(deny, ""), fails)
+    let denyj: String = containment_check_join(wt, "other-swarm")
+    let fails = ok("Rule 1: worker token may not JOIN another swarm", !str_eq(denyj, ""), fails)
+    let lat: String = containment_check_lateral(wt, "sibling-9")
+    let fails = ok("Rule 3: worker->worker lateral edge rejected", !str_eq(lat, ""), fails)
+    let ver: String = containment_check_lateral(wt, "")
+    let fails = ok("Rule 3: worker->manager vertical edge allowed", str_eq(ver, ""), fails)
+    // enforced live: a worker-token caller is denied opening a real swarm
+    let wcfg: String = json_set(cfg_r, "caller_token", wt)
+    let denied: String = swarm_run("analyze_item", refs, inputs, wcfg)
+    let fails = ok("Rule 2 enforced live: worker-caller swarm denied", str_eq(json_get_string(denied, "status"), "denied"), fails)
+
+    // ── E) vote convergence strategy at concurrency ──
+    let cfg_v: String = "{\"concurrency\":\"8\",\"strategy\":\"vote\",\"min_success_ratio\":\"1.0\"}"
+    let rv: String = swarm_run("classify", refs, inputs, cfg_v)
+    let winner: String = json_get_string(json_get_raw(rv, "merged"), "winner")
+    // billing/payments/ledger/invoicing/payroll/audit = long(>4); tax/fx = short -> long wins
+    let fails = ok("vote converged (winner=long)", str_eq(winner, "long"), fails)
+
+    // ── F) durable, inspectable work-tracking ──
+    let started: Int = worktrack_count_kind(corr, "worker.started")
+    let completed: Int = worktrack_count_kind(corr, "worker.completed")
+    let fails = ok("work-tracking journal: 8 started + 8 completed", (started == 8) && (completed == 8), fails)
+
+    print("")
+    if fails == 0 {
+        print("HARNESS GREEN — full local-swarm mechanics proven with seam=" + seam_mode())
+        return 0
+    }
+    print("HARNESS FAIL (" + int_to_str(fails) + ")")
+    return 1
+}

lang/swarm/tests/harness_real_cognition.el

@@ -0,0 +1,119 @@
+// harness_real_cognition.el — the LOCAL SWARM running REAL cognition.
+//
+// Run with: SWARM_PRIMITIVE_SEAM=decorated + the sandbox env sourced
+// (ENGRAM_URL=:8901). Each worker's `think` is BOUND to the reshape's proven
+// op_think (GET /api/think) over its NODE-ID anchor — real 768-dim gradients from
+// the live (isolated) geometry, not the stub. The @manager fans out N native-El
+// worker threads at real concurrency, converges (reduce + vote) over the real
+// cognition, enforces all three containment rules, observes afferent telemetry,
+// and work-tracks durably.
+//
+// Anchors are real self-neighbourhood node ids on the :8901 clone (free-text
+// anchors return "geometry unavailable", so these must be node ids).
+
+fn ok(label: String, cond: Bool, fails: Int) -> Int {
+    if cond { print("  ok    " + label); return fails }
+    print("  FAIL  " + label); return fails + 1
+}
+
+fn main() -> Int {
+    let fails = 0
+    print("== REAL-COGNITION LOCAL SWARM (seam=" + seam_mode() + ", engram=" + env("ENGRAM_URL") + ") ==")
+
+    // ── 0) direct proof the bound primitive returns REAL cognition ──
+    let g: String = op_think("self", "plan")
+    let dim: Int = json_get_int(g, "dim")
+    let nsup: Int = json_get_int(g, "n_support")
+    let fails = ok("bound op_think returns a real 768-dim gradient", dim == 768, fails)
+    let fails = ok("real gradient has support (n_support>0)", nsup > 0, fails)
+    let gfree: String = op_think("this-is-free-text-not-a-node", "reason")
+    let fails = ok("free-text anchor correctly refused (geometry unavailable)", str_contains(gfree, "geometry unavailable"), fails)
+
+    // ── the input set: 8 real NODE-ID anchors from self's neighbourhood ──
+    let anchors: String = "[\"a1000001-0000-0000-0000-000000000001\",\"5f011441-fa43-4fe7-a9c0-c78a584ef11d\",\"kn-5adecd7e-d6db-4576-87fe-6ef8a935cea6\",\"76d7fd0b-0672-4511-a2f5-a095cf9c60ae\",\"7027e302-593f-441d-8fd6-9c400c163108\",\"2a730b18-6566-46ee-a21e-4f4dd0380908\",\"46b0e4dd-2c19-48d2-bcbc-19f61d6c79ae\",\"9162cde8-8739-4f00-bfc9-2850ed612e50\"]"
+    let refs: String = "[\"self\"]"
+
+    // ── A) fan-out real cognition at concurrency, converge with REDUCE ──
+    let cfg_r: String = "{\"concurrency\":\"4\",\"strategy\":\"reduce\",\"min_success_ratio\":\"1.0\"}"
+    let rr: String = swarm_run("cognize", refs, anchors, cfg_r)
+    let fails = ok("swarm completed: 8 workers each a real think, concurrency=4", str_eq(json_get_string(rr, "status"), "completed"), fails)
+    let corr: String = json_get_string(rr, "corr_id")
+    let merged_r: String = json_get_raw(rr, "merged")
+    let fails = ok("reduce converged all 8 real-cognition outputs", str_to_int(json_get_string(merged_r, "count")) == 8, fails)
+    let acc: String = json_get_string(merged_r, "accumulated")
+    let fails = ok("converged output carries real gradient support (n_support)", str_contains(acc, "n_support"), fails)
+
+    // ── B) afferent telemetry: 8 real think-signals, decorated seam ──
+    let telem: String = json_get_raw(rr, "telemetry")
+    let aff: Int = str_to_int(json_get_string(telem, "afferent_think"))
+    let fails = ok("afferent counters = 8 real think invocations", aff == 8, fails)
+    let fails = ok("telemetry records seam_mode=decorated", str_eq(json_get_string(telem, "seam_mode"), "decorated"), fails)
+    let fails = ok("telemetry durably journalled", worktrack_count_kind(corr, "swarm.telemetry") == 1, fails)
+
+    // ── C) converge with VOTE over real cognition ──
+    let cfg_v: String = "{\"concurrency\":\"8\",\"strategy\":\"vote\",\"min_success_ratio\":\"1.0\"}"
+    let rv: String = swarm_run("cognize", refs, anchors, cfg_v)
+    let winner: String = json_get_string(json_get_raw(rv, "merged"), "winner")
+    let fails = ok("vote converged over real cognition (winner=" + winner + ")", !str_eq(winner, ""), fails)
+
+    // ── D) all three containment rules still enforced ──
+    let wt: String = containment_worker_token(corr, corr + "/worker-2")
+    let fails = ok("Rule 2: worker may not open a swarm", !str_eq(containment_check_open(wt), ""), fails)
+    let fails = ok("Rule 1: worker may not join another swarm", !str_eq(containment_check_join(wt, "s2"), ""), fails)
+    let fails = ok("Rule 3: worker->worker lateral edge rejected", !str_eq(containment_check_lateral(wt, "sib"), ""), fails)
+    let wcfg: String = json_set(cfg_r, "caller_token", wt)
+    let denied: String = swarm_run("cognize", refs, anchors, wcfg)
+    let fails = ok("Rule 2 enforced LIVE: worker-caller swarm denied", str_eq(json_get_string(denied, "status"), "denied"), fails)
+
+    // ── E) CCR scoping + non-leak over node-id anchors ──
+    let ctx: String = ccr_compile("cognize", refs, "a1000001-0000-0000-0000-000000000001", corr, corr + "/worker-0", wt)
+    let fails = ok("CCR context bounded within budget", ccr_within_budget(ctx), fails)
+    let leaks: Bool = str_contains(ctx, "9162cde8")
+    let fails = ok("CCR context does NOT leak sibling anchors", !leaks, fails)
+
+    // ── F) durable work-tracking ──
+    let started: Int = worktrack_count_kind(corr, "worker.started")
+    let completed: Int = worktrack_count_kind(corr, "worker.completed")
+    let fails = ok("work-tracking: 8 started + 8 completed", (started == 8) && (completed == 8), fails)
+
+    // ── G) RULE 4 — engram-write is @manager-ONLY (authority gate) ──
+    // A worker token (engram:read only) is STRUCTURALLY denied any engram write.
+    let worker_tok: String = containment_worker_token(corr, corr + "/worker-1")
+    let orch_tok: String = containment_coordinator_token(corr)
+    let fails = ok("worker token carries engram:read", containment_has_cap(worker_tok, "engram:read"), fails)
+    let fails = ok("worker token does NOT carry engram:write", !containment_has_cap(worker_tok, "engram:write"), fails)
+    let fails = ok("orchestrator token carries engram:write", containment_has_cap(orch_tok, "engram:write"), fails)
+    // a worker attempting an engram write is DENIED BY CAPABILITY (no HTTP issued)
+    let wdeny: String = swarm_engram_write(worker_tok, corr, "worker tries to mutate global state", "memory", 0.5)
+    let denied_reason: String = json_get_string(wdeny, "denied")
+    let fails = ok("worker engram-write DENIED by capability (Rule 4)", str_contains(denied_reason, "rule 4"), fails)
+    let fails = ok("denied worker write performed NO engram mutation (no node id)", str_eq(json_get_string(wdeny, "id"), ""), fails)
+    let fails = ok("Rule-4 violation journalled", worktrack_count_kind(corr, "containment.violation") >= 1, fails)
+    // the orchestrator passes the capability gate (sole authorized writer)
+    let odeny: String = containment_check_engram_write(orch_tok, "engram.write")
+    let fails = ok("orchestrator PASSES the engram-write capability gate (sole writer)", str_eq(odeny, ""), fails)
+
+    // ── H) curated merge = the only write path (orchestrator commits) ──
+    // The AUTHORITY gate above is already proven (worker denied, orchestrator
+    // authorized) WITHOUT issuing a write. The actual persisting commit exercises
+    // the engram write path, which needs the gate-1 write-healthy clone — so it
+    // runs only under SWARM_WRITE_HEALTHY=1 (else it would hit the known daemon
+    // write-crash). Authority != health: the gate holds either way.
+    if str_eq(env("SWARM_WRITE_HEALTHY"), "1") {
+        let cfg_commit: String = "{\"concurrency\":\"4\",\"strategy\":\"reduce\",\"min_success_ratio\":\"1.0\",\"commit\":\"1\"}"
+        let rc: String = swarm_run("cognize", refs, anchors, cfg_commit)
+        let committed: String = json_get_string(rc, "committed_node")
+        let fails2: Int = ok("orchestrator (sole writer) committed the merge to the engram", !str_eq(committed, ""), fails)
+        let fails = fails2
+    } else {
+        print("  note  curated-merge commit deferred to the gate-1 write-healthy clone (set SWARM_WRITE_HEALTHY=1); authority gate already proven above")
+    }
+
+    print("")
+    if fails == 0 {
+        print("REAL-COGNITION SWARM GREEN — Neuron thinking in parallel over its own geometry.")
+        return 0
+    }
+    print("REAL-COGNITION SWARM FAIL (" + int_to_str(fails) + ")")
+    return 1
+}

lang/swarm/tests/integ_engram.el

@@ -0,0 +1,45 @@
+// integ_engram.el — integration proof against a LIVE (isolated) engram.
+//
+// Run with the sandbox env sourced (ENGRAM_URL=http://127.0.0.1:8901,
+// ENGRAM_API_KEY=sbx-dev-swarm-ccr). Proves:
+//   (a) CCR retrieval pulls REAL content from the mind over HTTP;
+//   (b) a full swarm runs and converges against the live mind;
+//   (c) work-tracking mirrors records into the engram as SwarmTrack nodes.
+
+fn main() -> Int {
+    let url: String = env("ENGRAM_URL")
+    if str_eq(url, "") {
+        print("SKIP integ_engram (ENGRAM_URL not set)")
+        return 0
+    }
+
+    // (a) CCR compiles a bounded context whose retrieval hit the real mind.
+    let refs: String = "[\"Volatility-Based Decomposition\",\"Swarm Architecture containment\"]"
+    let wt: String = containment_worker_token("integ", "integ/w0")
+    let ctx: String = ccr_compile("analyze_item", refs, "decompose the billing module", "integ", "integ/w0", wt)
+    let knowledge: String = json_get_string(ctx, "knowledge")
+    let pulled_real: Bool = str_contains(knowledge, "olatility") || str_contains(knowledge, "Anderson") || str_contains(knowledge, "VBD")
+    if pulled_real {
+        print("  ok   CCR retrieval pulled real mind content (" + int_to_str(str_len(knowledge)) + " bytes, bounded)")
+    } else {
+        print("  FAIL CCR retrieval returned no mind content")
+    }
+    let bounded: Bool = ccr_within_budget(ctx)
+    if bounded { print("  ok   compiled context stayed within budget") } else { print("  FAIL context over budget") }
+
+    // (b) a real swarm over the live mind.
+    let inputs: String = "[\"billing\",\"payments\",\"ledger\"]"
+    let cfg: String = "{\"concurrency\":\"3\",\"strategy\":\"collect\",\"min_success_ratio\":\"1.0\"}"
+    let res: String = swarm_run("analyze_item", refs, inputs, cfg)
+    let status: String = json_get_string(res, "status")
+    if str_eq(status, "completed") { print("  ok   swarm completed against live engram") } else { print("  FAIL swarm status=" + status) }
+    let corr: String = json_get_string(res, "corr_id")
+
+    // (c) work-tracking mirrored into the mind: search for this swarm's records.
+    let hits: String = primitive_attend(corr, 5)
+    let mirrored: Bool = str_contains(hits, "swarm-track") || str_contains(hits, corr)
+    if mirrored { print("  ok   work-tracking mirrored into the engram (queryable)") } else { print("  note mirror not yet visible to search (async index)") }
+
+    print("DONE integ_engram corr=" + corr)
+    return 0
+}

lang/swarm/tests/test_convergence.el

@@ -0,0 +1,55 @@
+// test_convergence.el — convergence strategies + failure threshold / abort.
+
+fn assert_true(label: String, cond: Bool, fails: Int) -> Int {
+    if cond { print("  ok   " + label); return fails }
+    print("  FAIL " + label); return fails + 1
+}
+
+fn main() -> Int {
+    let fails = 0
+    let refs: String = "[]"
+
+    // ── vote: classify 5 inputs; 3 "long" (>4 chars) vs 2 "short" -> winner long ──
+    let inputs: String = "[\"alpha\",\"bravo\",\"hi\",\"charlie\",\"ok\"]"
+    let cfg_v: String = "{\"concurrency\":\"3\",\"strategy\":\"vote\",\"min_success_ratio\":\"1.0\"}"
+    let rv: String = swarm_run("classify", refs, inputs, cfg_v)
+    let merged_v: String = json_get_raw(rv, "merged")
+    let winner: String = json_get_string(merged_v, "winner")
+    let votes: Int = str_to_int(json_get_string(merged_v, "votes"))
+    let fails = assert_true("vote winner = long", str_eq(winner, "long"), fails)
+    let fails = assert_true("vote count = 3", votes == 3, fails)
+
+    // ── merge: outputs joined ──
+    let cfg_m: String = "{\"concurrency\":\"2\",\"strategy\":\"merge\",\"min_success_ratio\":\"1.0\"}"
+    let rm: String = swarm_run("analyze_item", refs, "[\"a\",\"b\",\"c\"]", cfg_m)
+    let merged_m: String = json_get_raw(rm, "merged")
+    let joined: String = json_get_string(merged_m, "merged")
+    let fails = assert_true("merge produced a joined string", str_contains(joined, "|"), fails)
+
+    // ── reduce: count accumulates ──
+    let cfg_r: String = "{\"concurrency\":\"4\",\"strategy\":\"reduce\",\"min_success_ratio\":\"1.0\"}"
+    let rr: String = swarm_run("analyze_item", refs, "[\"a\",\"b\",\"c\",\"d\"]", cfg_r)
+    let merged_r: String = json_get_raw(rr, "merged")
+    let rcount: Int = str_to_int(json_get_string(merged_r, "count"))
+    let fails = assert_true("reduce count = 4", rcount == 4, fails)
+
+    // ── failure threshold: 2 of 5 fail (x-prefixed); ratio 3/5=0.6 < 0.8 -> aborted ──
+    let fin: String = "[\"a\",\"xb\",\"c\",\"xd\",\"e\"]"
+    let cfg_f: String = "{\"concurrency\":\"5\",\"strategy\":\"collect\",\"min_success_ratio\":\"0.8\"}"
+    let rf: String = swarm_run("faildemo", refs, fin, cfg_f)
+    let fstatus: String = json_get_string(rf, "status")
+    let fails = assert_true("swarm aborted below min_success_ratio (0.6<0.8)", str_eq(fstatus, "aborted"), fails)
+    let corr_f: String = json_get_string(rf, "corr_id")
+    let failed_n: Int = worktrack_count_kind(corr_f, "worker.failed")
+    let aborted_n: Int = worktrack_count_kind(corr_f, "swarm.aborted")
+    let fails = assert_true("tracked 2 worker.failed", failed_n == 2, fails)
+    let fails = assert_true("tracked swarm.aborted", aborted_n == 1, fails)
+
+    // ── same failures tolerated when min_success_ratio=0.5 (0.6>=0.5) -> completed ──
+    let cfg_ok: String = "{\"concurrency\":\"5\",\"strategy\":\"collect\",\"min_success_ratio\":\"0.5\"}"
+    let rok: String = swarm_run("faildemo", refs, fin, cfg_ok)
+    let fails = assert_true("swarm completes when failures within tolerance", str_eq(json_get_string(rok, "status"), "completed"), fails)
+
+    if fails == 0 { print("PASS test_convergence"); return 0 }
+    print("FAIL test_convergence (" + int_to_str(fails) + ")"); return 1
+}

lang/swarm/tests/test_swarm.el

@@ -0,0 +1,75 @@
+// test_swarm.el — end-to-end proof of the swarm capability on native El threads.
+//
+// Proves: native-thread fan-out/converge, bounded concurrency, per-worker CCR
+// bounded context (with the security-boundary property), containment Rule 2
+// enforcement, and durable work-tracking.
+
+fn assert_true(label: String, cond: Bool, fails: Int) -> Int {
+    if cond {
+        print("  ok   " + label)
+        return fails
+    }
+    print("  FAIL " + label)
+    return fails + 1
+}
+
+fn main() -> Int {
+    let fails = 0
+
+    // ── 1) fan-out / converge (collect) over native threads ──
+    let inputs: String = "[\"alpha\",\"bravo\",\"charlie\",\"delta\",\"echo\"]"
+    let refs: String = "[]"
+    let cfg: String = "{\"concurrency\":\"2\",\"strategy\":\"collect\",\"min_success_ratio\":\"1.0\"}"
+    let res: String = swarm_run("analyze_item", refs, inputs, cfg)
+    let status: String = json_get_string(res, "status")
+    let fails = assert_true("swarm completed", str_eq(status, "completed"), fails)
+
+    let merged: String = json_get_raw(res, "merged")
+    let count: Int = json_array_len(merged)
+    let fails = assert_true("collect returned 5 results (bounded concurrency=2)", count == 5, fails)
+
+    // ── 2) work-tracking is durable + complete ──
+    let corr: String = json_get_string(res, "corr_id")
+    let started: Int = worktrack_count_kind(corr, "worker.started")
+    let completed: Int = worktrack_count_kind(corr, "worker.completed")
+    let created: Int = worktrack_count_kind(corr, "swarm.created")
+    let done: Int = worktrack_count_kind(corr, "swarm.completed")
+    let fails = assert_true("tracked 5 worker.started", started == 5, fails)
+    let fails = assert_true("tracked 5 worker.completed", completed == 5, fails)
+    let fails = assert_true("tracked swarm.created + swarm.completed", (created == 1) && (done == 1), fails)
+
+    // ── 3) CCR: bounded, minimal, non-leaking per-worker context ──
+    let wtoken: String = containment_worker_token(corr, corr + "/worker-0")
+    let ctx: String = ccr_compile("analyze_item", refs, "alpha", corr, corr + "/worker-0", wtoken)
+    let in_budget: Bool = ccr_within_budget(ctx)
+    let fails = assert_true("CCR context within token budget", in_budget, fails)
+    let this_input: String = json_get_string(ctx, "input")
+    let fails = assert_true("CCR context contains THIS worker's input", str_eq(this_input, "alpha"), fails)
+    // security boundary: a worker's compiled context must not carry a sibling input
+    let leaks_sibling: Bool = str_contains(ctx, "charlie")
+    let fails = assert_true("CCR context does NOT leak sibling inputs", !leaks_sibling, fails)
+
+    // ── 4) containment Rule 2: a worker may not open a swarm ──
+    let worker_caller_cfg: String = json_set(cfg, "caller_token", wtoken)
+    let denied: String = swarm_run("analyze_item", refs, inputs, worker_caller_cfg)
+    let dstatus: String = json_get_string(denied, "status")
+    let fails = assert_true("worker-token caller denied opening a swarm (Rule 2)", str_eq(dstatus, "denied"), fails)
+
+    // coordinator token IS allowed
+    let coord: String = containment_coordinator_token("some-corr")
+    let allow_reason: String = containment_check_open(coord)
+    let fails = assert_true("coordinator token allowed to open a swarm", str_eq(allow_reason, ""), fails)
+
+    // ── 5) containment Rule 3: no lateral worker->worker edge ──
+    let lateral: String = containment_check_lateral(wtoken, "some-sibling")
+    let fails = assert_true("lateral worker->worker edge rejected (Rule 3)", !str_eq(lateral, ""), fails)
+    let vertical: String = containment_check_lateral(wtoken, "")
+    let fails = assert_true("vertical worker->coordinator edge allowed", str_eq(vertical, ""), fails)
+
+    if fails == 0 {
+        print("PASS test_swarm")
+        return 0
+    }
+    print("FAIL test_swarm (" + int_to_str(fails) + " failures)")
+    return 1
+}

lang/swarm/tests/test_worktrack.el

@@ -0,0 +1,38 @@
+// test_worktrack.el — durability + inspectability of the work-tracking journal.
+
+fn main() -> Int {
+    let corr: String = "test-" + uuid_v4()
+
+    // record a swarm lifecycle
+    let p1: String = json_set("{}", "input_count", "3")
+    worktrack_append("swarm.created", corr, "swarm-1", p1)
+    worktrack_append("worker.started", corr, "worker-001", "{}")
+    worktrack_append("worker.started", corr, "worker-002", "{}")
+    worktrack_append("worker.completed", corr, "worker-001", "{}")
+    worktrack_append("worker.failed", corr, "worker-002", "{}")
+    worktrack_append("swarm.completed", corr, "swarm-1", "{}")
+
+    // inspect: reconstruct the report from the durable journal
+    let report: String = worktrack_swarm_report(corr)
+    print("report=" + report)
+
+    let recs_n: Int = el_list_len(worktrack_records(corr))
+    print("records=" + int_to_str(recs_n))
+
+    let state: String = json_get_string(report, "state")
+    let completed: Int = str_to_int(json_get_string(report, "workers_completed"))
+    let failed: Int = str_to_int(json_get_string(report, "workers_failed"))
+
+    if str_eq(state, "completed") {
+        if completed == 1 {
+            if failed == 1 {
+                if recs_n == 6 {
+                    print("PASS worktrack")
+                    return 0
+                }
+            }
+        }
+    }
+    print("FAIL worktrack")
+    return 1
+}

lang/swarm/worktrack.el

@@ -0,0 +1,217 @@
+// worktrack.el — full work-tracking for the swarm.
+//
+// "Intent all the way up, orchestrator at the top." Every unit of parallel
+// work a swarm fans out is recorded here: the swarm itself, each worker, its
+// status, its result summary, the convergence, and the final merged output —
+// all threaded by a single correlation ID so the entire execution graph can be
+// reconstructed and audited (Swarm Architecture §6.1).
+//
+// DURABILITY. Records are appended to a JSON-lines journal on disk. The journal
+// is append-only and single-writer: only the coordinator (the main thread, before
+// and after each fan-out and during convergence) writes to it. Workers never
+// touch it — they return structured results and the coordinator records them.
+// This is deliberate: it makes the tracking store race-free and, not
+// coincidentally, enforces Swarm containment rule 3 (no lateral worker state).
+//
+// INSPECTABILITY. The journal is plain JSONL — greppable, tailable, replayable.
+// worktrack_read() loads it back; worktrack_swarm_report() reconstructs a
+// swarm's full record from its correlation ID.
+//
+// ENGRAM MIRROR (optional). When ENGRAM_URL is set, each record is also mirrored
+// into the engram as a node (POST /api/node) tagged with the correlation ID, so
+// the swarm's execution becomes part of the durable mind, queryable by memory.
+//
+// Depends on: el_runtime.c builtins (fs_*, http_post, env, json_*, uuid_v4,
+// now_millis, str_*). No El-module concat dependencies of its own.
+
+// ── JSON helper ──────────────────────────────────────────────────────────────
+// json_set inserts its value as a RAW JSON fragment (objects/arrays/numbers).
+// json_set_str sets a plain STRING value, correctly quoted and escaped. Use
+// json_set for nested JSON, json_set_str for strings.
+fn json_set_str(j: String, key: String, val: String) -> String {
+    return json_set(j, key, "\"" + json_escape_string(val) + "\"")
+}
+
+// ── Journal location ─────────────────────────────────────────────────────────
+
+// worktrack_dir — directory holding the swarm journals.
+//   Override with SWARM_TRACK_DIR; defaults to ./.swarm-track (relative to CWD).
+fn worktrack_dir() -> String {
+    let d: String = env("SWARM_TRACK_DIR")
+    if str_eq(d, "") {
+        return ".swarm-track"
+    }
+    return d
+}
+
+// worktrack_journal_path — the JSONL journal file for one correlation ID.
+fn worktrack_journal_path(corr_id: String) -> String {
+    return worktrack_dir() + "/" + corr_id + ".jsonl"
+}
+
+// worktrack_init — ensure the journal directory exists. Idempotent.
+fn worktrack_init() -> Bool {
+    let d: String = worktrack_dir()
+    if fs_exists(d) {
+        return true
+    }
+    return fs_mkdir(d)
+}
+
+// ── Record construction ──────────────────────────────────────────────────────
+
+// worktrack_record — build one journal record as a JSON object string.
+//   kind:    the record kind (swarm.created, worker.started, ...)
+//   corr_id: the swarm correlation ID (links every record)
+//   subject: the entity the record is about (swarm id, worker id, "")
+//   payload: a JSON object string with kind-specific fields
+fn worktrack_record(kind: String, corr_id: String, subject: String, payload: String) -> String {
+    let kv: [String] = el_list_empty()
+    let kv = el_list_append(kv, "kind")
+    let kv = el_list_append(kv, kind)
+    let kv = el_list_append(kv, "corr_id")
+    let kv = el_list_append(kv, corr_id)
+    let kv = el_list_append(kv, "subject")
+    let kv = el_list_append(kv, subject)
+    let kv = el_list_append(kv, "ts_ms")
+    let kv = el_list_append(kv, int_to_str(now_millis()))
+    let rec: String = json_build_object(kv)
+    // Attach the payload as a nested raw JSON field.
+    let rec2: String = json_set(rec, "data", payload)
+    return rec2
+}
+
+// ── Journal append (single-writer, durable) ──────────────────────────────────
+
+// worktrack_append — append one record to the correlation journal (durable),
+// and mirror it to the engram if ENGRAM_URL is configured. Returns the record.
+//
+// fs_write here is used in append semantics: we read-modify-write the file. The
+// coordinator is the only writer, so this is safe and race-free.
+fn worktrack_append(kind: String, corr_id: String, subject: String, payload: String) -> String {
+    worktrack_init()
+    let rec: String = worktrack_record(kind, corr_id, subject, payload)
+    let path: String = worktrack_journal_path(corr_id)
+    let prior: String = ""
+    if fs_exists(path) {
+        let prior = fs_read(path)
+    }
+    let next: String = prior + rec + "\n"
+    fs_write(path, next)
+    worktrack_mirror_engram(rec, corr_id, kind, subject)
+    return rec
+}
+
+// worktrack_mirror_engram — best-effort mirror of a record into the engram.
+// No-op unless ENGRAM_URL is set. Failures are swallowed (tracking must not
+// depend on the mind being reachable).
+fn worktrack_mirror_engram(rec: String, corr_id: String, kind: String, subject: String) -> Bool {
+    // Opt-in: the durable substrate is the JSONL journal (always written). The
+    // engram mirror is an additional convenience, enabled with SWARM_MIRROR=1,
+    // so a swarm never depends on — or loads — the mind just to track its work.
+    if str_eq(env("SWARM_MIRROR"), "1") {
+        // enabled — fall through to the mirror POST
+        let _go: Int = 1
+    } else {
+        return false
+    }
+    let url: String = env("ENGRAM_URL")
+    if str_eq(url, "") {
+        return false
+    }
+    let content: String = "swarm-track " + kind + " " + subject + " :: " + rec
+    let body_kv: [String] = el_list_empty()
+    let body_kv = el_list_append(body_kv, "content")
+    let body_kv = el_list_append(body_kv, content)
+    let body_kv = el_list_append(body_kv, "node_type")
+    let body_kv = el_list_append(body_kv, "SwarmTrack")
+    let body_kv = el_list_append(body_kv, "salience")
+    let body_kv = el_list_append(body_kv, "0.5")
+    let body: String = json_build_object(body_kv)
+    let key: String = env("ENGRAM_API_KEY")
+    let body2: String = json_set_str(body, "_auth", key)
+    let resp: String = http_post(url + "/api/nodes", body2)
+    return true
+}
+
+// ── Read / inspect ───────────────────────────────────────────────────────────
+
+// worktrack_read — read the raw JSONL journal for a correlation ID.
+fn worktrack_read(corr_id: String) -> String {
+    let path: String = worktrack_journal_path(corr_id)
+    if fs_exists(path) {
+        return fs_read(path)
+    }
+    return ""
+}
+
+// worktrack_records — the journal as a [String] of record JSON objects, in order.
+fn worktrack_records(corr_id: String) -> [String] {
+    let raw: String = worktrack_read(corr_id)
+    let out: [String] = el_list_empty()
+    if str_eq(raw, "") {
+        return out
+    }
+    let lines: [String] = str_split_lines(raw)
+    let n: Int = el_list_len(lines)
+    let i = 0
+    while i < n {
+        let ln: String = el_list_get(lines, i)
+        if str_eq(ln, "") {
+            let i = i + 1
+        } else {
+            let out = el_list_append(out, ln)
+            let i = i + 1
+        }
+    }
+    return out
+}
+
+// worktrack_count_kind — how many records of a given kind exist for a swarm.
+// Powers assertions and live status ("how many workers completed").
+fn worktrack_count_kind(corr_id: String, kind: String) -> Int {
+    let recs: [String] = worktrack_records(corr_id)
+    let n: Int = el_list_len(recs)
+    let c = 0
+    let i = 0
+    while i < n {
+        let r: String = el_list_get(recs, i)
+        let k: String = json_get_string(r, "kind")
+        if str_eq(k, kind) {
+            let c = c + 1
+        }
+        let i = i + 1
+    }
+    return c
+}
+
+// worktrack_swarm_report — reconstruct a compact status report for a swarm from
+// its journal: counts of started/completed/failed workers and terminal state.
+// Inspectable, durable, derived purely from the append-only record.
+fn worktrack_swarm_report(corr_id: String) -> String {
+    let started: Int = worktrack_count_kind(corr_id, "worker.started")
+    let completed: Int = worktrack_count_kind(corr_id, "worker.completed")
+    let failed: Int = worktrack_count_kind(corr_id, "worker.failed")
+    let done: Int = worktrack_count_kind(corr_id, "swarm.completed")
+    let aborted: Int = worktrack_count_kind(corr_id, "swarm.aborted")
+    let state: String = "running"
+    if aborted > 0 {
+        let state = "aborted"
+    } else {
+        if done > 0 {
+            let state = "completed"
+        }
+    }
+    let kv: [String] = el_list_empty()
+    let kv = el_list_append(kv, "corr_id")
+    let kv = el_list_append(kv, corr_id)
+    let kv = el_list_append(kv, "state")
+    let kv = el_list_append(kv, state)
+    let kv = el_list_append(kv, "workers_started")
+    let kv = el_list_append(kv, int_to_str(started))
+    let kv = el_list_append(kv, "workers_completed")
+    let kv = el_list_append(kv, int_to_str(completed))
+    let kv = el_list_append(kv, "workers_failed")
+    let kv = el_list_append(kv, int_to_str(failed))
+    return json_build_object(kv)
+}

lang/tests/bench/fitprobe.el

@@ -0,0 +1,91 @@
+// fitprobe.el — controlled growth-curve specimens for validating the complexity fitter.
+//
+// Three deliberately-shaped workloads. None depends on a real defect existing,
+// which is the point: the fitter must be provable against KNOWN curves.
+//
+//   linear   — one allocation per item.        count O(n),  bytes O(n),   time O(n)
+//   accum    — rebuilds its accumulator.       count O(n),  bytes O(n^2), time O(n^2)
+//   compute  — nested arithmetic, no alloc.    count O(1),  bytes O(1),   time O(n^2)
+//
+// `compute` is the specimen that matters. It is the shape of el #132
+// (strlen-per-character inside str_char_code): pure CPU, zero allocation.
+// An allocation-only gate is structurally blind to it.
+//
+// No imports — uses runtime builtins directly so nothing collides.
+
+fn work_linear(n: Int) -> Int {
+    let parts: [String] = native_list_empty()
+    let i: Int = 0
+    while i < n {
+        let parts = native_list_append(parts, int_to_str(i))
+        let i = i + 1
+    }
+    return native_list_len(parts)
+}
+
+fn work_accum(n: Int) -> Int {
+    let acc: String = ""
+    let i: Int = 0
+    while i < n {
+        let acc = acc + "x"
+        let i = i + 1
+    }
+    return str_len(acc)
+}
+
+fn work_compute(n: Int) -> Int {
+    // str_char_code is an opaque external call, so the C optimiser cannot
+    // reduce this nest to a closed form the way it does with `total + 1`.
+    // This is the exact shape of el #132: n scans over n characters, pure
+    // CPU, ZERO allocation.
+    let s: String = "abcdefghij"
+    let total: Int = 0
+    let i: Int = 0
+    while i < n {
+        let j: Int = 0
+        while j < n {
+            let total = total + str_char_code(s, 0)
+            let j = j + 1
+        }
+        let i = i + 1
+    }
+    return total
+}
+
+fn run_one(mode: String, n: Int) {
+    let c0: Int = el_alloc_count()
+    let b0: Int = el_alloc_bytes()
+    let t0: Int = el_now_instant()
+
+    let r: Int = 0
+    if str_eq(mode, "linear")  { let r = work_linear(n) }
+    if str_eq(mode, "accum")   { let r = work_accum(n) }
+    if str_eq(mode, "compute") { let r = work_compute(n) }
+
+    let t1: Int = el_now_instant()
+    let c1: Int = el_alloc_count()
+    let b1: Int = el_alloc_bytes()
+
+    println(mode + "\t" + int_to_str(n)
+        + "\t" + int_to_str(c1 - c0)
+        + "\t" + int_to_str(b1 - b0)
+        + "\t" + int_to_str((t1 - t0) / 1000)
+        + "\t" + int_to_str(r))
+    return
+}
+
+fn sweep(mode: String) {
+    run_one(mode, 200)
+    run_one(mode, 400)
+    run_one(mode, 800)
+    run_one(mode, 1600)
+    return
+}
+
+fn main() -> Int {
+    println("mode\tn\tallocs\tbytes\tusec\tsink")
+    sweep("linear")
+    sweep("accum")
+    sweep("compute")
+    return 0
+}

lang/tests/native/test_compiler.el

@@ -1,3 +1,4 @@
+import "../../runtime/eltest.el"
 // tests/native/test_compiler.el — comprehensive tests for the El compiler pipeline.
 //
 // Tests the lexer (lexer.el), parser (parser.el), and codegen (codegen.el)

lang/tests/native/test_core.el

@@ -1,3 +1,4 @@
+import "../../runtime/eltest.el"
 // test_codegen_js.el - basic tests for JS codegen features.
 //
 // These tests verify that core El language features produce correct values

lang/tests/native/test_elbench.el

@@ -0,0 +1,111 @@
+import "../../runtime/eltest.el"
+import "../../runtime/elbench.el"
+
+// test_elbench.el — proves the growth-curve classifier against KNOWN curves.
+//
+// Every series below is real measured data from lang/tests/bench/fitprobe.el
+// on a geometric sweep n = 200/400/800/1600. The classifier must be provable
+// without depending on a live defect existing, which is the whole point of
+// keeping controlled specimens.
+
+fn _s4(a: Int, b: Int, c: Int, d: Int) -> [Int] {
+    let l: [Int] = native_list_empty()
+    let l = native_list_append(l, a)
+    let l = native_list_append(l, b)
+    let l = native_list_append(l, c)
+    let l = native_list_append(l, d)
+    return l
+}
+
+test "classifies a linear allocation series as O(n)" {
+    // fitprobe `linear`, allocation count
+    let v = _s4(208, 409, 810, 1611)
+    assert elb_measured_curve(v, 10) == 2, "linear allocs should classify O(n)"
+}
+
+test "classifies a linear byte series as O(n)" {
+    // fitprobe `linear`, allocation bytes
+    let v = _s4(4786, 9682, 19474, 39658)
+    assert elb_measured_curve(v, 10) == 2, "linear bytes should classify O(n)"
+}
+
+test "classifies a quadratic byte series as O(n^2)" {
+    // fitprobe `accum`, allocation bytes -- the accumulator-rebuild shape
+    let v = _s4(20300, 80600, 321200, 1282400)
+    assert elb_measured_curve(v, 10) == 4, "accum bytes should classify O(n^2)"
+}
+
+test "accumulator count is linear -- proves count alone misses it" {
+    // Same run as above. The COUNT is exactly linear while bytes are
+    // quadratic. A count-only gate passes this defect clean.
+    let v = _s4(200, 400, 800, 1600)
+    assert elb_measured_curve(v, 10) == 2, "accum count classifies O(n)"
+    assert elb_gate(v, 2, 10) == 0, "count-only gate PASSES the quadratic"
+}
+
+test "classifies a quadratic time series as O(n^2)" {
+    // fitprobe `compute` -- el #132's shape: n scans over n characters
+    let v = _s4(67, 205, 818, 3268)
+    assert elb_measured_curve(v, 10) == 4, "compute time should classify O(n^2)"
+}
+
+test "REFUSES an all-zero series instead of calling it O(1)" {
+    // fitprobe `compute` allocation count. Pure CPU, allocates nothing.
+    // Reporting O(1) here would be a confident answer with nothing behind it.
+    let v = _s4(0, 0, 0, 0)
+    assert elb_gate(v, 2, 10) == 3, "all-zero series must be REFUSED"
+    assert elb_measured_curve(v, 10) < 0, "unclassifiable returns -1"
+}
+
+test "REFUSES an implausibly flat series" {
+    // The shape produced when clang closes a loop to a multiply: a real
+    // answer, no work done, no movement across an 8x input range.
+    let v = _s4(1000, 1001, 1002, 1003)
+    assert elb_gate(v, 2, 10) == 3, "hard-flat series must be REFUSED"
+}
+
+test "gate FAILS a quadratic declared as linear" {
+    let v = _s4(20300, 80600, 321200, 1282400)
+    assert elb_gate(v, 2, 10) == 1, "O(n^2) measured vs O(n) declared must FAIL"
+}
+
+test "gate PASSES a linear series declared as linear" {
+    let v = _s4(208, 409, 810, 1611)
+    assert elb_gate(v, 2, 10) == 0, "O(n) measured vs O(n) declared must PASS"
+}
+
+test "gate reports BETTER when measured beats the declared bound" {
+    let v = _s4(208, 409, 810, 1611)
+    assert elb_gate(v, 4, 10) == 4, "O(n) measured vs O(n^2) declared is BETTER"
+}
+
+test "gate reports INDETERMINATE on disagreeing ratios" {
+    // fitprobe `linear` WALL TIME at these sizes: 26/19/43/78 microseconds.
+    // Ratios 0.73, 2.26, 1.81 disagree well past the noise threshold. The
+    // honest answer is "cannot tell", not a classification -- this is exactly
+    // why benchmarks need auto-scaled iteration counts rather than one shot.
+    let v = _s4(26, 19, 43, 78)
+    assert elb_gate(v, 2, 10) == 2, "disagreeing ratios must be INDETERMINATE"
+}
+
+test "black_box is a real barrier and returns its input" {
+    assert el_black_box(42) == 42, "black_box is value-preserving"
+    let s: Int = 0
+    let i: Int = 0
+    while i < 100 {
+        // Bind the call before using it in arithmetic: `x + call(...)`
+        // lowers to el_str_concat() on integers. Same inference defect
+        // as `call(...) == y` lowering to str_eq().
+        let bx: Int = el_black_box(1)
+        let s = s + bx
+        let i = i + 1
+    }
+    assert s == 100, "black_box does not disturb the computation"
+}
+
+test "curve names round-trip" {
+    assert elb_curve_from_name("O(n)") == 2, "O(n) parses"
+    assert elb_curve_from_name("O(n^2)") == 4, "O(n^2) parses"
+    assert str_eq(elb_curve_name(4), "O(n^2)"), "O(n^2) renders"
+    assert elb_curve_from_name("O(nonsense)") < 0, "unknown curve is -1"
+}

lang/tests/native/test_env.el

@@ -1,3 +1,4 @@
+import "../../runtime/eltest.el"
 // test_env.el - native test suite for runtime/env.el
 //
 // Covers: env() for reading environment variables, args() returning a list,

lang/tests/native/test_fs.el

@@ -1,3 +1,4 @@
+import "../../runtime/eltest.el"
 // test_fs.el - native test suite for runtime/fs.el
 //
 // Covers: fs_write/read round-trip, fs_exists, fs_mkdir, fs_list,

lang/tests/native/test_json.el

@@ -1,3 +1,4 @@
+import "../../runtime/eltest.el"
 // test_json.el - native test suite for runtime/json.el
 //
 // Covers: json_get (dot-path), typed extractors (int, bool, float),

lang/tests/native/test_lexer_scaling.el

@@ -0,0 +1,178 @@
+import "../../runtime/eltest.el"
+import "../../runtime/elbench.el"
+
+// test_lexer_scaling.el — THE ARMED GATE.
+//
+// This is the regression test that would have caught el #132.
+//
+// #132 was a strlen() inside str_char_code() and str_slice(). The lexer walks
+// source one character at a time, so every character access rescanned the whole
+// remaining input: O(n) per character over n characters = O(n^2). It shipped for
+// months. It was found by a geometric sweep, not by reading code.
+//
+// So this test IS a geometric sweep. It scans a string of length n, character by
+// character, at four doubling sizes, and asserts the cost is linear. If anyone
+// reintroduces a per-character rescan — in str_char_code, in str_slice, in any
+// accessor the lexer leans on — the measured curve becomes O(n^2) and this fails.
+//
+// The value is in it being ARMED, not in it currently failing. It passes today
+// because #132 is fixed. That is the correct state for a regression gate.
+//
+// Note the deliberate `let c: Int = str_char_code(...)` binding in the scan loop.
+// Inlining it as `total + str_char_code(s, i)` lowers to el_str_concat() on
+// integers — the Plus arm of the operator-typing family, still open at the time
+// of writing. Binding first is the safe form.
+
+// _mk_string — build a string of length >= n by DOUBLING.
+//
+// Deliberately not `s = s + "x"` n times: that is itself quadratic in bytes and
+// would contaminate the very measurement this test exists to take. Doubling
+// allocates ~2n total.
+fn _mk_string(n: Int) -> String {
+    let s: String = "abcdefgh"
+    while str_len(s) < n {
+        let s = s + s
+    }
+    return s
+}
+
+// _scan — walk the string one character at a time, REPS times.
+//
+// This is the lexer's access pattern reduced to its essential shape. The
+// repetitions lift the measurement clear of timer resolution; without them the
+// smaller sizes land in noise and the classifier correctly reports
+// INDETERMINATE rather than guessing.
+fn _scan(s: String, n: Int, reps: Int) -> Int {
+    let total: Int = 0
+    let r: Int = 0
+    while r < reps {
+        let i: Int = 0
+        while i < n {
+            let c: Int = str_char_code(s, i)
+            let total = total + c
+            let i = i + 1
+        }
+        let r = r + 1
+    }
+    return total
+}
+
+// _measure_scan — microseconds for a full scan sweep point.
+fn _measure_scan(n: Int, reps: Int) -> Int {
+    let s: String = _mk_string(n)
+    // WARMUP, discarded. Without it the small-n end of the sweep is dominated
+    // by cold caches and reads as superlinear on genuinely linear work --
+    // measured ratios 3.37 2.92 1.76 1.65 on exactly this workload.
+    let w: Int = _scan(s, n, 2)
+    let wj: Int = el_black_box(w)
+    let t0: Int = el_now_instant()
+    let got: Int = _scan(s, n, reps)
+    let t1: Int = el_now_instant()
+    // Feed the result through the barrier so the scan cannot be elided.
+    let sink: Int = el_black_box(got)
+    if sink == 0 { println("") }
+    return (t1 - t0) / 1000
+}
+
+fn _series4(a: Int, b: Int, c: Int, d: Int) -> [Int] {
+    let l: [Int] = native_list_empty()
+    let l = native_list_append(l, a)
+    let l = native_list_append(l, b)
+    let l = native_list_append(l, c)
+    let l = native_list_append(l, d)
+    return l
+}
+
+test "character scan is LINEAR in time -- regression gate for el #132" {
+    let reps: Int = 40
+    let t1: Int = _measure_scan(16384, reps)
+    let t2: Int = _measure_scan(32768, reps)
+    let t3: Int = _measure_scan(65536, reps)
+    let t4: Int = _measure_scan(131072, reps)
+    let series: [Int] = _series4(t1, t2, t3, t4)
+
+    let verdict: Int = elb_gate(series, 2, 50)
+    let measured: Int = elb_measured_curve(series, 50)
+
+    // Report the actual numbers regardless of outcome. A gate that fires
+    // without showing its evidence is just an assertion.
+    println("      scan us: " + int_to_str(t1) + " " + int_to_str(t2) + " "
+        + int_to_str(t3) + " " + int_to_str(t4)
+        + "  -> " + elb_curve_name(measured) + " [" + elb_verdict_name(verdict) + "]")
+
+    // PASS (0) or BETTER (4) are both acceptable. FAIL (1) means someone
+    // reintroduced superlinear per-character cost. REFUSED (3) or
+    // INDETERMINATE (2) mean the measurement is untrustworthy -- which is
+    // also a failure of this test, deliberately: a gate that cannot measure
+    // must not report success.
+    assert verdict == 0 || verdict == 4, "character scan must measure O(n) or better"
+}
+
+test "string building by doubling stays linear in allocated bytes" {
+    let b1: Int = el_alloc_bytes()
+    let s1: String = _mk_string(8192)
+    let b2: Int = el_alloc_bytes()
+    let s2: String = _mk_string(16384)
+    let b3: Int = el_alloc_bytes()
+    let s3: String = _mk_string(32768)
+    let b4: Int = el_alloc_bytes()
+    let s4: String = _mk_string(65536)
+    let b5: Int = el_alloc_bytes()
+
+    let series: [Int] = _series4(b2 - b1, b3 - b2, b4 - b3, b5 - b4)
+    let verdict: Int = elb_gate(series, 2, 1000)
+    let measured: Int = elb_measured_curve(series, 1000)
+    println("      bytes: " + int_to_str(b2 - b1) + " " + int_to_str(b3 - b2) + " "
+        + int_to_str(b4 - b3) + " " + int_to_str(b5 - b4)
+        + "  -> " + elb_curve_name(measured) + " [" + elb_verdict_name(verdict) + "]")
+
+    assert verdict == 0 || verdict == 4, "doubling build must be O(n) in bytes"
+    assert str_len(s4) >= 65536, "final string reached the requested size"
+}
+
+// _scan_quadratic — a DELIBERATELY quadratic scan: for each position, rescan
+// from the start. This is precisely what el #132 did — strlen() from offset 0
+// on every character access — reproduced here so the gate can be proven to
+// FIRE, not merely to pass on healthy code. An unproven gate is decoration.
+fn _scan_quadratic(s: String, n: Int) -> Int {
+    let total: Int = 0
+    let i: Int = 0
+    while i < n {
+        let j: Int = 0
+        while j < i {
+            let c: Int = str_char_code(s, j)
+            let total = total + c
+            let j = j + 1
+        }
+        let i = i + 1
+    }
+    return total
+}
+
+fn _measure_quadratic(n: Int) -> Int {
+    let s: String = _mk_string(n)
+    let w: Int = _scan_quadratic(s, 64)
+    let wj: Int = el_black_box(w)
+    let t0: Int = el_now_instant()
+    let got: Int = _scan_quadratic(s, n)
+    let t1: Int = el_now_instant()
+    let sink: Int = el_black_box(got)
+    return (t1 - t0) / 1000
+}
+
+test "the gate FIRES on a live quadratic scan -- proves it is armed" {
+    let q1: Int = _measure_quadratic(1024)
+    let q2: Int = _measure_quadratic(2048)
+    let q3: Int = _measure_quadratic(4096)
+    let q4: Int = _measure_quadratic(8192)
+    let series: [Int] = _series4(q1, q2, q3, q4)
+
+    let verdict: Int = elb_gate(series, 2, 50)
+    let measured: Int = elb_measured_curve(series, 50)
+    println("      quad us: " + int_to_str(q1) + " " + int_to_str(q2) + " "
+        + int_to_str(q3) + " " + int_to_str(q4)
+        + "  -> " + elb_curve_name(measured) + " [" + elb_verdict_name(verdict) + "]")
+
+    assert measured == 4, "a rescan-from-zero workload must classify O(n^2)"
+    assert verdict == 1, "declared O(n) against measured O(n^2) must FAIL the gate"
+}

lang/tests/native/test_math.el

@@ -1,3 +1,4 @@
+import "../../runtime/eltest.el"
 // test_math.el - native test suite for runtime/math.el
 //
 // Covers: integer math (abs, max, min), float math (sqrt, log, sin, cos, pi),

lang/tests/native/test_state.el

@@ -1,3 +1,4 @@
+import "../../runtime/eltest.el"
 // test_state.el - native test suite for runtime/state.el
 //
 // Covers: state_set/get/del, state_has, state_get_or, state_keys,

lang/tests/native/test_string.el

@@ -1,3 +1,4 @@
+import "../../runtime/eltest.el"
 // test_string.el - native test suite for runtime/string.el
 //
 // Covers: type conversions, core primitives, comparison and search,

lang/tests/native/test_text.el

@@ -1,3 +1,4 @@
+import "../../runtime/eltest.el"
 // test_text.el - native test suite for text primitives.
 //
 // Mirrors the acceptance corpus in tests/text/examples/ using the

lang/tests/native/test_time.el

@@ -1,3 +1,4 @@
+import "../../runtime/eltest.el"
 // test_time.el - native test suite for runtime/time.el
 //
 // Covers: time_now (positive timestamp), time_to_parts (UTC decomposition),

lang/tests/native/test_transduce.el

@@ -0,0 +1,234 @@
+import "../../runtime/eltest.el"
+// test_transduce.el — geometry as a first-class El value, and realizers
+// declared in El rather than patched into the runtime.
+//
+// WHAT IS ACTUALLY UNDER TEST. Until 2026-08-16 no El ingest path could carry
+// a vector: nodes took text, and geometry was DERIVED from that text. Text was
+// therefore the mandatory entry medium, so any non-text modality had to be
+// DESCRIBED in prose first and the geometry we reasoned over was the geometry
+// OF THE DESCRIPTION, not of the signal. The fix has two halves, and this file
+// exercises both:
+//
+//   1. Geometry is a VALUE — it carries its own width, so nothing has to
+//      assert a width against a string's length.
+//   2. A REALIZER is an ordinary El function. `tone_realizer` below is not in
+//      the runtime, is not known to the compiler, and is not special in any
+//      way; it is registered BY NAME and dispatched to through transduce().
+//      That is the load-bearing claim: adding a modality must not require a
+//      runtime patch, or nothing has actually moved into the language.
+//
+// COMPARISON DISCIPLINE IN THIS FILE (measured 2026-08-16, not stylistic):
+// elc lowers `a == b` to a NUMERIC comparison only when both operand names are
+// in the per-function int-name set, which `let x: Int` populates. A bare call
+// like `geometry_is(g) == 0` is not a registered name, so it lowers to
+// `str_eq(...)` — strcmp on two integers reinterpreted as pointers. `<` and `>`
+// lower directly via binop_to_c with no type inference at all, so truthiness is
+// written `> 0` / `< 1` here, and any exact `==` is done on a value first bound
+// through `let x: Int`.
+
+// ── A realizer, written entirely in El ──────────────────────────────────────
+// Maps a "tone" signal into a 4-component geometry. Deliberately trivial —
+// what is being proven is that an El function can BE a realizer, not that
+// this is good acoustics. The one real property it has: distinct signals
+// produce distinct geometry, so the test can tell transduction from a stub.
+fn tone_realizer(signal: String) -> Geometry {
+    let g: Geometry = geometry_new(4)
+    let n: Int = str_len(signal)
+    let a: Int = geometry_set(g, 0, int_to_float(n))
+    let b: Int = geometry_set(g, 1, int_to_float(n * 2))
+    let c: Int = geometry_set(g, 2, int_to_float(n * 3))
+    let d: Int = geometry_set(g, 3, int_to_float(n * 4))
+    g
+}
+
+// A second realizer for a different modality, to prove the registry keys on
+// modality and does not just hand back "the last thing registered".
+fn pulse_realizer(signal: String) -> Geometry {
+    let g: Geometry = geometry_new(2)
+    let a: Int = geometry_set(g, 0, 1.0)
+    let b: Int = geometry_set(g, 1, 0.0)
+    g
+}
+
+// A deliberately BROKEN realizer: it returns something that is not a Geometry.
+// transduce() must not hand this back to a caller as if it were one.
+fn bogus_realizer(signal: String) -> Geometry {
+    return 12345
+}
+
+test "geometry-is-a-value-with-its-own-width" {
+    let g: Geometry = geometry_new(8)
+    let live: Int = geometry_is(g)
+    assert live > 0, "geometry_new returns a live Geometry"
+    let d: Int = geometry_dim(g)
+    assert d == 8, "a Geometry carries its own width"
+    let freed: Int = geometry_free(g)
+    assert freed > 0, "geometry_free reports what it did"
+}
+
+test "geometry-rejects-nonsense-without-an-arbitrary-bound" {
+    // dim <= 0 is not a width. Note there is deliberately no MAX dim here:
+    // #141 needed `dim <= 8192` only to bound an allocation sized from a
+    // caller's claim about a string. A value that carries its own width has
+    // nothing left to validate, so the only failure left is allocation.
+    let zero: Geometry = geometry_new(0)
+    let z: Int = geometry_is(zero)
+    assert z < 1, "dim 0 is not a geometry"
+    let neg: Geometry = geometry_new(-4)
+    let n: Int = geometry_is(neg)
+    assert n < 1, "negative dim is not a geometry"
+    // Accessors must be total: a non-geometry is 0-width, never a crash.
+    let nd: Int = geometry_dim(0)
+    assert nd < 1, "geometry_dim of a non-geometry is 0"
+    let ni: Int = geometry_is(0)
+    assert ni < 1, "geometry_is of a non-geometry is 0"
+    let nf: Int = geometry_free(0)
+    assert nf < 1, "geometry_free of a non-geometry is a no-op"
+}
+
+test "geometry-components-round-trip" {
+    let g: Geometry = geometry_new(3)
+    let s0: Int = geometry_set(g, 0, 1.5)
+    let s1: Int = geometry_set(g, 1, -2.5)
+    assert s0 > 0, "set in range succeeds"
+    let oob: Int = geometry_set(g, 3, 9.0)
+    assert oob < 1, "set out of range is refused, not silently dropped"
+    let v0: Float = geometry_get(g, 0)
+    let d0: Float = v0 - 1.5
+    assert d0 < 0.001, "component 0 round-trips"
+    assert d0 > -0.001, "component 0 round-trips"
+    let v1: Float = geometry_get(g, 1)
+    let d1: Float = v1 + 2.5
+    assert d1 < 0.001, "component 1 round-trips (negative)"
+    assert d1 > -0.001, "component 1 round-trips (negative)"
+    let freed: Int = geometry_free(g)
+}
+
+test "hex-is-an-edge-adapter-and-derives-its-own-width" {
+    // 2 components, little-endian float32: 1.0 = 0000803f, 2.0 = 00000040.
+    let g: Geometry = geometry_from_f32le_hex("0000803f00000040")
+    let live: Int = geometry_is(g)
+    assert live > 0, "valid hex decodes to a Geometry"
+    let d: Int = geometry_dim(g)
+    assert d == 2, "width is DERIVED from the input, never supplied"
+    let a: Float = geometry_get(g, 0)
+    let da: Float = a - 1.0
+    assert da < 0.001, "first component decoded"
+    assert da > -0.001, "first component decoded"
+    let b: Float = geometry_get(g, 1)
+    let db: Float = b - 2.0
+    assert db < 0.001, "second component decoded"
+    assert db > -0.001, "second component decoded"
+    // Egress adapter is the exact inverse.
+    let back: String = geometry_to_f32le_hex(g)
+    assert str_eq(back, "0000803f00000040"), "hex round-trips exactly"
+    let freed: Int = geometry_free(g)
+}
+
+test "hex-rejects-malformed-input" {
+    let empty: Geometry = geometry_from_f32le_hex("")
+    let e: Int = geometry_is(empty)
+    assert e < 1, "empty hex is not a geometry"
+    let ragged: Geometry = geometry_from_f32le_hex("0000803f0000")
+    let r: Int = geometry_is(ragged)
+    assert r < 1, "length not a multiple of 8 is refused"
+    let nonhex: Geometry = geometry_from_f32le_hex("zzzzzzzz")
+    let nh: Int = geometry_is(nonhex)
+    assert nh < 1, "non-hex characters are refused"
+}
+
+test "a-realizer-declared-in-el-is-a-first-class-realizer" {
+    // THE CLAIM: tone_realizer is an ordinary El function. It is not in the
+    // runtime and the compiler knows nothing about it. Registering it by name
+    // is enough to make it the organ for a modality.
+    let reg: Int = realizer_register("tone", "tone_realizer")
+    assert reg > 0, "an El fn registers as a realizer by name"
+    let has: Int = realizer_has("tone")
+    assert has > 0, "the modality now has an organ"
+
+    let g: Geometry = transduce("aaa", "tone")
+    let live: Int = geometry_is(g)
+    assert live > 0, "transduce returns real geometry"
+    let d: Int = geometry_dim(g)
+    assert d == 4, "the El realizer determined the width, not the runtime"
+    // str_len("aaa") == 3, so component 0 must be 3.0 — proof the signal
+    // actually reached the El function rather than a stub answering for it.
+    let c0: Float = geometry_get(g, 0)
+    let dc: Float = c0 - 3.0
+    assert dc < 0.001, "the signal reached the El realizer"
+    assert dc > -0.001, "the signal reached the El realizer"
+    let freed: Int = geometry_free(g)
+}
+
+test "distinct-signals-transduce-to-distinct-geometry" {
+    let reg: Int = realizer_register("tone", "tone_realizer")
+    let g1: Geometry = transduce("aa", "tone")
+    let g2: Geometry = transduce("aaaaa", "tone")
+    let a: Float = geometry_get(g1, 0)
+    let b: Float = geometry_get(g2, 0)
+    let diff: Float = b - a
+    // 5 - 2 = 3. If transduction were a stub these would be equal.
+    assert diff > 2.9, "different signals produce different geometry"
+    assert diff < 3.1, "different signals produce different geometry"
+    let f1: Int = geometry_free(g1)
+    let f2: Int = geometry_free(g2)
+}
+
+test "the-registry-keys-on-modality" {
+    let r1: Int = realizer_register("tone", "tone_realizer")
+    let r2: Int = realizer_register("pulse", "pulse_realizer")
+    assert r2 > 0, "a second modality registers independently"
+    let gt: Geometry = transduce("aaa", "tone")
+    let gp: Geometry = transduce("aaa", "pulse")
+    let dt: Int = geometry_dim(gt)
+    let dp: Int = geometry_dim(gp)
+    assert dt == 4, "tone still routes to its own realizer"
+    assert dp == 2, "pulse routes to a different realizer"
+    let f1: Int = geometry_free(gt)
+    let f2: Int = geometry_free(gp)
+}
+
+test "no-organ-is-reported-as-no-organ" {
+    // A modality with no realizer must transduce to NOTHING. It must never
+    // fall back to embedding a description of the signal and calling that
+    // perception — that silent substitution is the entire defect this change
+    // exists to end.
+    let has: Int = realizer_has("echolocation")
+    assert has < 1, "unregistered modality has no organ"
+    let g: Geometry = transduce("anything", "echolocation")
+    let live: Int = geometry_is(g)
+    assert live < 1, "no realizer means no geometry, not fake geometry"
+}
+
+test "registration-of-an-unresolvable-name-fails-loudly" {
+    // Reported at the moment of WIRING, not later as "this modality mysteriously
+    // produces nothing". Distinguishing "no organ" from "broken organ" is the
+    // lesson that made this whole change necessary.
+    let bad: Int = realizer_register("ghost", "no_such_function_anywhere")
+    assert bad < 1, "an unresolvable realizer name is a registration failure"
+    let has: Int = realizer_has("ghost")
+    assert has < 1, "and nothing gets registered"
+}
+
+test "a-realizer-returning-non-geometry-transduces-nothing" {
+    let reg: Int = realizer_register("bogus", "bogus_realizer")
+    assert reg > 0, "the symbol resolves, so registration succeeds"
+    // ...but the contract is enforced at the boundary, so the caller never
+    // receives a value that would misbehave far away from here.
+    let g: Geometry = transduce("x", "bogus")
+    let live: Int = geometry_is(g)
+    assert live < 1, "a non-Geometry return transduced nothing"
+}
+
+test "norm-lets-a-caller-check-a-realizer-emitted-signal" {
+    let g: Geometry = geometry_new(2)
+    let z: Float = geometry_norm(g)
+    assert z < 0.001, "a fresh geometry is zero — norm says so"
+    let s0: Int = geometry_set(g, 0, 3.0)
+    let s1: Int = geometry_set(g, 1, 4.0)
+    let n: Float = geometry_norm(g)
+    let dn: Float = n - 5.0
+    assert dn < 0.001, "3-4-5: norm is 5"
+    assert dn > -0.001, "3-4-5: norm is 5"
+    let freed: Int = geometry_free(g)
+}

lang/tests/runtime/operator_typing_test.el

@@ -0,0 +1,28 @@
+fn getstr(x: String) -> String { return x }
+fn getint(x: Int) -> Int { return x }
+fn ok(label: String) -> Void { println("ok   " + label) }
+fn bad(label: String) -> Void { println("FAIL " + label) }
+
+let s1: String = "hello"
+let s2: String = "hello"
+let s3: String = "world"
+let i1: Int = 5
+let i2: Int = 5
+let i3: Int = 9
+
+if "abc" == "abc" { ok("str literal eq") } else { bad("str literal eq") }
+if "abc" == "xyz" { bad("str literal ne") } else { ok("str literal ne") }
+if s1 == s2 { ok("str var eq") } else { bad("str var eq") }
+if s1 == s3 { bad("str var ne") } else { ok("str var ne") }
+if getstr("hi") == "hi" { ok("str call vs literal") } else { bad("str call vs literal") }
+if s1 == getstr("hello") { ok("str var vs call") } else { bad("str var vs call") }
+if s1 == getstr("nope") { bad("str var vs call ne") } else { ok("str var vs call ne") }
+if i1 == i2 { ok("int var eq") } else { bad("int var eq") }
+if i1 == i3 { bad("int var ne") } else { ok("int var ne") }
+if getint(5) == i1 { ok("int call vs var") } else { bad("int call vs var") }
+if getint(9) == i1 { bad("int call vs var ne") } else { ok("int call vs var ne") }
+if s1 != s3 { ok("str NOTEQ") } else { bad("str NOTEQ") }
+if s1 != s2 { bad("str NOTEQ same") } else { ok("str NOTEQ same") }
+if i1 != i3 { ok("int NOTEQ") } else { bad("int NOTEQ") }
+if getint(9) != i1 { ok("int call NOTEQ") } else { bad("int call NOTEQ") }
+println("done")

lang/tests/runtime/str_cache_test.el

@@ -0,0 +1,58 @@
+fn expect_int(label: String, got: Int, want: Int) -> Void {
+    if got == want { println("ok   " + label) }
+    else { println("FAIL " + label + " got=" + int_to_str(got) + " want=" + int_to_str(want)) }
+}
+fn expect_str(label: String, got: String, want: String) -> Void {
+    if str_eq(got, want) { println("ok   " + label) }
+    else { println("FAIL " + label + " got='" + got + "' want='" + want + "'") }
+}
+
+// 1. basic char access across a string
+let s: String = "hello"
+expect_int("char[0]=h", str_char_code(s, 0), 104)
+expect_int("char[4]=o", str_char_code(s, 4), 111)
+expect_int("char[5] OOB -> 0", str_char_code(s, 5), 0)
+expect_int("char[-1] OOB -> 0", str_char_code(s, -1), 0)
+expect_int("empty string OOB", str_char_code("", 0), 0)
+
+// 2. slices
+expect_str("slice(0,5)", str_slice(s, 0, 5), "hello")
+expect_str("slice(1,3)", str_slice(s, 1, 3), "el")
+expect_str("slice past end clamps", str_slice(s, 3, 99), "lo")
+expect_str("slice inverted -> empty", str_slice(s, 4, 2), "")
+
+// 3. DIFFERENT strings must not share a cached length (the real hazard)
+let a: String = "abc"
+let b: String = "abcdefghij"
+expect_int("a[2]=c", str_char_code(a, 2), 99)
+expect_int("a[3] OOB", str_char_code(a, 3), 0)
+expect_int("b[9]=j", str_char_code(b, 9), 106)
+expect_int("b[3]=d after a", str_char_code(b, 3), 100)
+expect_int("a[3] still OOB after b", str_char_code(a, 3), 0)
+
+// 4. many distinct strings interleaved — forces cache slot collisions
+fn interleave(n: Int) -> Int {
+    let i: Int = 0
+    let bad: Int = 0
+    while i < n {
+        let t: String = int_to_str(i)
+        let l: Int = str_len(t)
+        let last: Int = str_char_code(t, l - 1)
+        let oob: Int = str_char_code(t, l)
+        if oob != 0 { let bad2: Int = bad + 1
+                      let bad: Int = bad2 }
+        if last == 0 { let bad3: Int = bad + 1
+                       let bad: Int = bad3 }
+        let i2: Int = i + 1
+        let i: Int = i2
+    }
+    return bad
+}
+expect_int("1000 interleaved strings, no bad reads", interleave(1000), 0)
+
+// 5. concatenation changes length — cache must not report the old one
+let g: String = "12345"
+let g2: String = g + "6789"
+expect_int("grown string len via char", str_char_code(g2, 8), 57)
+expect_int("original still bounded", str_char_code(g, 5), 0)
+println("done")

lang/tests/runtime/string_test.el

@@ -1,3 +1,4 @@
+import "../../runtime/eltest.el"
 // tests/runtime/string_test.el — Test suite for runtime/string.el
 //
 // Exercises every public function exported by runtime/string.el using the

tools/neuron-sandbox/README.md

@@ -173,4 +173,7 @@ Each ad-hoc harness becomes `nsbx run <name> …` (or `--source` build) against
 ## Env knobs
 
 `NSBX_ROOT`, `NSBX_PORT_BASE`, `NSBX_RSS_BOUND_MB`, `NSBX_REMERGE_THRESHOLD`,
+`NSBX_READY_TIMEOUT_SECS` (default 15 — how long `up`/`create`/`build` wait for a
+daemon to answer `/api/stats` before reporting failure; raise it if a boot is
+legitimately slow under concurrent sandbox/CPU load rather than actually broken),
 `EL_REPO` (for `elc` + runtime sources), `ENGRAM_LIVE_DATA_DIR`, `ENGRAM_LIVE_PLIST`.

tools/neuron-sandbox/nsbx

@@ -32,6 +32,12 @@ EL_REPO="${EL_REPO:-$HOME/Development/neuron-technologies/foundation/el}"
 PORT_BASE="${NSBX_PORT_BASE:-8900}"
 RSS_BOUND_MB="${NSBX_RSS_BOUND_MB:-550}"      # from store-fix reboot-proof (aaf13f88)
 REMERGE_THRESHOLD="${NSBX_REMERGE_THRESHOLD:-40000}"
+# readiness-poll window for start_daemon (0.5s ticks). Default unchanged (15s) —
+# but a cold boot against the full live store, under concurrent CPU contention
+# from other running sandboxes, has been observed live to take well past that.
+# Bump per-invocation with NSBX_READY_TIMEOUT_SECS if `up`/`create` reports a
+# not-ready failure but the daemon looks otherwise fine (see its logs/daemon.log).
+READY_TICKS=$(( ${NSBX_READY_TIMEOUT_SECS:-15} * 2 ))
 KEYSTONES=( "kn-efeb4a5b-5aff-4759-8a97-7233099be6ee" "kn-5b606390-a52d-4ca2-8e0e-eba141d13440" )
 # fixed probe set for retrieval-parity (stable, identity-anchored)
 PARITY_QUERIES=( "who am I" "self identity core" "engram store durability" "keystone self anchor" "grounding honesty" )
@@ -77,7 +83,7 @@ _port_claimed(){ # is another sandbox already assigned this port?
 live_stats(){ curl -s -m5 "$LIVE_URL/api/stats" 2>/dev/null; }
 api(){ # api <name> <path> [json-body]
   local name="$1" path="$2" body="${3:-}"
-  local port; port="$(mget "$name" "['port']")"; [ -n "$port" ] || die "unknown sandbox: $name"
+  local port; port="$(mget "$name" "['port']")"; [ -n "$port" ] || die "unknown sandbox: $name (run: nsbx list)"
   local url="http://127.0.0.1:${port}${path}"
   if [ -n "$body" ]; then curl -s -m30 -X POST -H 'Content-Type: application/json' -d "$body" "$url"
   else curl -s -m30 "$url"; fi
@@ -88,6 +94,19 @@ stat_field(){ printf '%s' "$1" | sed -n "s/.*\"$2\":\([0-9]*\).*/\1/p"; }
 daemon_pid(){ local f; f="$(sdir "$1")/daemon.pid"; [ -f "$f" ] && cat "$f" || true; }
 daemon_alive(){ local p; p="$(daemon_pid "$1")"; [ -n "$p" ] && kill -0 "$p" 2>/dev/null; }
 
+# daemon_health <name> : prints "stopped" | "running" | "unresponsive" (to stdout).
+# "running" means the pid is alive AND /api/stats actually answered — process
+# liveness alone (daemon_alive) is not proof the HTTP server is serving; a pegged
+# or hung process still passes kill -0. Short timeout (2s) since this runs per-row
+# in `nsbx list`.
+daemon_health(){
+  local name="$1"
+  daemon_alive "$name" || { echo "stopped"; return 0; }
+  local port; port="$(mget "$name" "['port']")"
+  local s; s="$(curl -s -m2 "http://127.0.0.1:${port}/api/stats" 2>/dev/null)"
+  [ -n "$s" ] && echo "running" || echo "unresponsive"
+}
+
 # ---------------------------------------------------------------- elc/build ----
 find_elc(){
   command -v elc 2>/dev/null && return 0
@@ -123,6 +142,49 @@ _build_binary(){
   ok "built: $out ($(ls -lh "$out" | awk '{print $5}'), sha $(sha "$out" | cut -c1-12))"
 }
 
+# bin_built_at <path> : human-readable build timestamp. `cp -p` preserves mtime,
+# so this is the ORIGINAL build time even for binaries copied stock-prod into a
+# sandbox — not the copy time.
+bin_built_at(){ stat -f '%Sm' -t '%Y-%m-%d %H:%M:%S' "$1" 2>/dev/null || echo "unknown"; }
+
+# _binary_freshness <name> : best-effort staleness note (empty string if fresh/
+# unknown — never guesses). Two cases:
+#  - stock-prod: compares the sha recorded at create time against the CURRENTLY
+#    configured live real binary's sha (recomputed now, not cached) — catches
+#    "live prod moved on since this sandbox was cloned".
+#  - branch/source/rebuilt: compares the recorded source_commit against the
+#    LOCAL origin/dev ref (no fetch — reads whatever the repo already has) —
+#    catches "built from a commit that predates current dev tip".
+_binary_freshness(){
+  local name="$1" src; src="$(mget "$name" "['source']")"
+  case "$src" in
+    stock-prod:*)
+      local live_bin cur_sha rec_sha
+      live_bin="$(_live_real_bin)"
+      [ -n "$live_bin" ] && [ -x "$live_bin" ] || return 0
+      cur_sha="$(sha "$live_bin")"; rec_sha="$(mget "$name" "['binary_sha256']")"
+      [ -n "$cur_sha" ] && [ -n "$rec_sha" ] && [ "$cur_sha" != "$rec_sha" ] \
+        && printf 'stale: live prod binary has moved on since this sandbox was cloned (live is now %s, sha %s) — nsbx build %s --binary %s to catch up' \
+             "$(basename "$live_bin")" "${cur_sha:0:12}" "$name" "$live_bin"
+      ;;
+    branch:*|source:*|rebuilt:*)
+      local commit cur behind
+      commit="$(mget "$name" "['source_commit']")"
+      [ -n "$commit" ] || return 0
+      cur="$(git -C "$EL_REPO" rev-parse origin/dev 2>/dev/null)" || return 0
+      [ -n "$cur" ] && [ "$commit" != "$cur" ] || return 0
+      git -C "$EL_REPO" merge-base --is-ancestor "$commit" "$cur" 2>/dev/null || return 0
+      behind="$(git -C "$EL_REPO" rev-list --count "$commit..$cur" 2>/dev/null)"
+      printf 'stale: built from %s, %s commit(s) behind local origin/dev (%s) — nsbx build %s --branch origin/dev' \
+        "${commit:0:12}" "${behind:-?}" "${cur:0:12}" "$name"
+      ;;
+  esac
+}
+
+# _source_commit <dir> : best-effort git HEAD of a source tree used to build a
+# sandbox binary, empty if not a git repo (e.g. a prebuilt --binary path has none).
+_source_commit(){ git -C "$1" rev-parse HEAD 2>/dev/null || true; }
+
 # ---------------------------------------------------------------- daemon -------
 # start_daemon <name> : boots the sandbox's real engram binary on its isolated
 # port against its cloned data dir, with the SAME auto-remerge net the live soul
@@ -133,9 +195,9 @@ start_daemon(){
   local port bin data export key
   port="$(mget "$name" "['port']")"; bin="$d/bin/engram"; data="$d/data"
   key="sbx-$name"; export="$data/.scan-export.reseed-clean.json"
-  [ -x "$bin" ] || die "sandbox binary missing: $bin"
+  [ -x "$bin" ] || die "sandbox binary missing: $bin (run: nsbx build $name --source DIR | --branch REF, or nsbx destroy $name && nsbx create $name to reclone stock-prod)"
   [ "$port" != "$LIVE_BIND_PORT" ] && [ "$port" != "$SOUL_PORT" ] || die "refusing forbidden port $port"
-  [ -f "$data/neuron.egm" ] || die "sandbox has no cloned store: $data/neuron.egm"
+  [ -f "$data/neuron.egm" ] || die "sandbox has no cloned store: $data/neuron.egm (data dir is corrupt/incomplete — run: nsbx destroy $name && nsbx create $name)"
   # HARD guard: never point a sandbox daemon at the live data dir.
   [ "$(cd "$data" && pwd -P)" != "$(cd "$LIVE_DATA_DIR" && pwd -P)" ] || die "refusing: sandbox data dir resolves to LIVE store"
 
@@ -150,11 +212,11 @@ start_daemon(){
   echo "$pid" > "$d/daemon.pid"
   # readiness poll
   local url="http://127.0.0.1:$port" i s
-  for i in $(seq 1 30); do
+  for i in $(seq 1 "$READY_TICKS"); do
     s="$(curl -s -m3 "$url/api/stats" 2>/dev/null)"
     [ -n "$s" ] && break; sleep 0.5
   done
-  [ -n "$s" ] || { warn "daemon did not become ready (see $d/logs/daemon.log)"; return 1; }
+  [ -n "$s" ] || { warn "daemon did not become ready within ${NSBX_READY_TIMEOUT_SECS:-15}s (see $d/logs/daemon.log). pid $pid may still be alive and slow to boot under load — check: lsof -iTCP:$port -P, or retry with NSBX_READY_TIMEOUT_SECS=45"; return 1; }
   ok "ready pid=$pid  boot-stats: $s"
   # auto-remerge net (idempotent): match live edge population if the export is present
   if [ -f "$export" ]; then
@@ -231,33 +293,35 @@ cmd_create(){
   info "live baseline stats: ${lstats:-<unavailable>}"
 
   # ---- determine + place the runtime binary (versioned into the snapshot) ----
-  local source_desc live_bin
+  local source_desc live_bin source_commit=""
   live_bin="$(_live_real_bin)"
   if [ -n "$binpath" ]; then
     [ -x "$binpath" ] || die "not an executable binary: $binpath"
     cp -p "$binpath" "$d/bin/engram"; source_desc="prebuilt:$binpath"
   elif [ -n "$src" ]; then
     _build_binary "$src" "$d/bin/engram" "$d/build"; source_desc="source:$src"
+    source_commit="$(_source_commit "$src")"
   elif [ -n "$branch" ]; then
     log "worktree: $repo @ $branch -> $d/build/worktree"
     git -C "$repo" worktree add --detach "$d/build/worktree" "$branch" >/dev/null 2>&1 \
       || die "git worktree add failed ($repo @ $branch)"
     _build_binary "$d/build/worktree" "$d/bin/engram" "$d/build"; source_desc="branch:$branch@$repo"
+    source_commit="$(_source_commit "$d/build/worktree")"
   else
     [ -x "$live_bin" ] || die "cannot resolve live ENGRAM_REAL_BIN: $live_bin"
     cp -p "$live_bin" "$d/bin/engram"; source_desc="stock-prod:$live_bin"
   fi
   local bin_sha; bin_sha="$(sha "$d/bin/engram")"
-  info "runtime: $source_desc  (sha ${bin_sha:0:12})"
+  info "runtime: $source_desc  (sha ${bin_sha:0:12}, built $(bin_built_at "$d/bin/engram"))"
 
   # ---- write manifest ----
-  python3 - "$name" "$port" "$source_desc" "$bin_sha" "$egm_sha" "$base_nodes" "$base_edges" "$(sha "$live_bin" 2>/dev/null)" <<'PY' > "$(manifest "$name")"
+  python3 - "$name" "$port" "$source_desc" "$bin_sha" "$egm_sha" "$base_nodes" "$base_edges" "$(sha "$live_bin" 2>/dev/null)" "$source_commit" <<'PY' > "$(manifest "$name")"
 import json,sys,datetime
-name,port,src,binsha,egmsha,bn,be,livebinsha=sys.argv[1:9]
+name,port,src,binsha,egmsha,bn,be,livebinsha,source_commit=sys.argv[1:10]
 json.dump({
  "name":name,"port":int(port),"created_at":datetime.datetime.now(datetime.timezone.utc).isoformat(),
  "source":src,"binary_sha256":binsha,"clone_egm_sha256":egmsha,
- "live_binary_sha256":livebinsha,
+ "live_binary_sha256":livebinsha,"source_commit":source_commit,
  "live_baseline":{"node_count":int(bn or 0),"edge_count":int(be or 0)},
  "keystones":["kn-efeb4a5b-5aff-4759-8a97-7233099be6ee","kn-5b606390-a52d-4ca2-8e0e-eba141d13440"]
 }, sys.stdout, indent=2)
@@ -268,6 +332,17 @@ PY
   start_daemon "$name" || die "daemon failed to start"
   local sstats; sstats="$(sbx_stats "$name")"
   local sbn sbe; sbn="$(stat_field "$sstats" node_count)"; sbe="$(stat_field "$sstats" edge_count)"
+  # a boot immediately followed by an auto-remerge can leave the daemon briefly
+  # busy — retry rather than silently folding a 0/0 baseline into the manifest.
+  # `validate`'s zero-loss/reboot-prove checks compare current counts >= baseline,
+  # so a 0/0 baseline would make them trivially PASS regardless of real data loss.
+  local _bi
+  for _bi in 1 2 3 4 5; do
+    [ -n "$sbn" ] && [ "$sbn" != "0" ] && break
+    sleep 1
+    sstats="$(sbx_stats "$name")"; sbn="$(stat_field "$sstats" node_count)"; sbe="$(stat_field "$sstats" edge_count)"
+  done
+  [ -z "$sbn" ] || [ "$sbn" = "0" ] && warn "sandbox stats still empty/zero after retries — recording sbx_baseline 0/0. This makes 'nsbx validate $name' zero-loss checks trivially pass; investigate before trusting a validate PASS: nsbx status $name"
   _capture_retrieval "$name" "$d/baseline/retrieval.json"
   # fold sandbox baseline into manifest
   python3 - "$(manifest "$name")" "$sbn" "$sbe" <<'PY'
@@ -320,7 +395,12 @@ except Exception: print("[]")' 2>/dev/null)"
 # thereafter. Prod on :$LIVE_BIND_PORT/:$SOUL_PORT is unreachable from here by design.
 cmd_up(){
   local name; if [ $# -gt 0 ] && [ "${1#-}" = "$1" ]; then name="$1"; shift; else name="${USER:-dev}-dev"; fi
-  if mexists "$name"; then daemon_alive "$name" || start_daemon "$name"; else cmd_create "$name" "$@"; fi
+  if mexists "$name"; then
+    daemon_alive "$name" || start_daemon "$name" \
+      || die "daemon did not become ready — see $(sdir "$name")/logs/daemon.log (try: nsbx up $name  again once you've checked the log)"
+  else
+    cmd_create "$name" "$@"
+  fi
   local port; port="$(mget "$name" "['port']")"
   echo >&2
   ok "your sandbox '$name' is ready at http://127.0.0.1:$port  (a private copy of the mind — prod is untouchable)"
@@ -334,34 +414,37 @@ cmd_up(){
 # it on the SAME clone + port (the code-change dev loop, in place).
 cmd_build(){
   local name="$1"; shift || true
-  mexists "$name" || die "no such sandbox: $name"
+  mexists "$name" || die "no such sandbox: $name (run: nsbx list — or nsbx create $name to make it)"
   local src="" branch="" repo="$EL_REPO"
   while [ $# -gt 0 ]; do case "$1" in
     --source) src="$2"; shift 2;; --branch) branch="$2"; shift 2;; --repo) repo="$2"; shift 2;;
     *) die "unknown flag: $1";; esac; done
   local d; d="$(sdir "$name")"
   stop_daemon "$name"
-  if [ -n "$src" ]; then _build_binary "$src" "$d/bin/engram" "$d/build"
+  local source_commit=""
+  if [ -n "$src" ]; then _build_binary "$src" "$d/bin/engram" "$d/build"; source_commit="$(_source_commit "$src")"
   elif [ -n "$branch" ]; then
     rm -rf "$d/build/worktree" 2>/dev/null; git -C "$repo" worktree prune 2>/dev/null
     git -C "$repo" worktree add --detach "$d/build/worktree" "$branch" >/dev/null 2>&1 || die "worktree add failed"
     _build_binary "$d/build/worktree" "$d/bin/engram" "$d/build"
+    source_commit="$(_source_commit "$d/build/worktree")"
   else die "usage: nsbx build <name> --source DIR | --branch REF [--repo R]"; fi
   # record new binary sha
-  python3 - "$(manifest "$name")" "$(sha "$d/bin/engram")" "${src:-branch:$branch}" <<'PY'
-import json,sys; mf,s,src=sys.argv[1:4]
-d=json.load(open(mf)); d["binary_sha256"]=s; d["source"]="rebuilt:"+src
+  python3 - "$(manifest "$name")" "$(sha "$d/bin/engram")" "${src:-branch:$branch}" "$source_commit" <<'PY'
+import json,sys; mf,s,src,source_commit=sys.argv[1:5]
+d=json.load(open(mf)); d["binary_sha256"]=s; d["source"]="rebuilt:"+src; d["source_commit"]=source_commit
 json.dump(d,open(mf,'w'),indent=2)
 PY
-  start_daemon "$name"
+  start_daemon "$name" \
+    || die "rebuilt binary did not become ready — see $d/logs/daemon.log (the old binary is gone; fix the code and re-run nsbx build $name ...)"
   ok "rebuilt + restarted on :$(mget "$name" "['port']")"
 }
 
 # ================================================================ run ==========
 cmd_run(){
   local name="$1"; shift || true
-  mexists "$name" || die "no such sandbox: $name"
-  daemon_alive "$name" || start_daemon "$name"
+  mexists "$name" || die "no such sandbox: $name (run: nsbx list — or nsbx create $name to make it)"
+  daemon_alive "$name" || start_daemon "$name" || die "daemon not running and failed to start — see $(sdir "$name")/logs/daemon.log"
   local d port; d="$(sdir "$name")"; port="$(mget "$name" "['port']")"
   # direct API form: nsbx run <name> api <path> [json]
   if [ "${1:-}" = "api" ]; then
@@ -395,8 +478,8 @@ cmd_run(){
 # RSS bound; retrieval parity; keystone integrity.
 cmd_validate(){
   local name="$1"; shift || true
-  mexists "$name" || die "no such sandbox: $name"
-  daemon_alive "$name" || start_daemon "$name"
+  mexists "$name" || die "no such sandbox: $name (run: nsbx list — or nsbx create $name to make it)"
+  daemon_alive "$name" || start_daemon "$name" || die "daemon not running and failed to start — see $(sdir "$name")/logs/daemon.log"
   local d port key; d="$(sdir "$name")"; port="$(mget "$name" "['port']")"; key="sbx-$name"
   local url="http://127.0.0.1:$port"
   local bn be; bn="$(mget "$name" "['sbx_baseline']['node_count']")"; be="$(mget "$name" "['sbx_baseline']['edge_count']")"
@@ -489,7 +572,7 @@ PY
 # Default is a DRY-RUN plan; requires --i-approve-prod-cutover to actually cut over.
 cmd_promote(){
   local name="$1"; shift || true
-  mexists "$name" || die "no such sandbox: $name"
+  mexists "$name" || die "no such sandbox: $name (run: nsbx list — or nsbx create $name to make it)"
   local approve=0 do_data=0
   while [ $# -gt 0 ]; do case "$1" in
     --i-approve-prod-cutover) approve=1; shift;;
@@ -588,7 +671,7 @@ PY
 # ================================================================ destroy ======
 cmd_destroy(){
   local name="$1"; shift || true
-  mexists "$name" || die "no such sandbox: $name"
+  mexists "$name" || die "no such sandbox: $name (run: nsbx list — or nsbx create $name to make it)"
   local d; d="$(sdir "$name")"
   stop_daemon "$name"
   if [ -d "$d/build/worktree" ]; then
@@ -604,22 +687,44 @@ cmd_destroy(){
 # ================================================================ list/status ==
 cmd_list(){
   [ -d "$SBX_ROOT" ] || { echo "no sandboxes"; return 0; }
-  printf '%-16s %-6s %-8s %-9s %s\n' NAME PORT STATE PID SOURCE
+  printf '%-16s %-6s %-13s %-9s %-19s %s\n' NAME PORT STATE PID "BUILT" SOURCE
   local m
   for m in "$SBX_ROOT"/*/manifest.json; do
     [ -f "$m" ] || continue
-    local n p src pid state
+    local n p src pid state bpath built fresh
     n="$(python3 -c "import json;print(json.load(open('$m'))['name'])")"
     p="$(python3 -c "import json;print(json.load(open('$m'))['port'])")"
     src="$(python3 -c "import json;print(json.load(open('$m'))['source'])")"
-    pid="$(daemon_pid "$n")"; state="stopped"; daemon_alive "$n" && state="running"
-    printf '%-16s %-6s %-8s %-9s %s\n' "$n" "$p" "$state" "${pid:-–}" "$src"
+    pid="$(daemon_pid "$n")"
+    case "$(daemon_health "$n")" in
+      running)      state="running";;
+      unresponsive) state="running(!resp)";;
+      *)            state="stopped";;
+    esac
+    bpath="$(sdir "$n")/bin/engram"; built="$([ -f "$bpath" ] && bin_built_at "$bpath" || echo unknown)"
+    fresh="$(_binary_freshness "$n")"; [ -n "$fresh" ] && src="[STALE] $src"
+    printf '%-16s %-6s %-13s %-9s %-19s %s\n' "$n" "$p" "$state" "${pid:-–}" "$built" "$src"
   done
+  info "state 'running(!resp)' = process alive but /api/stats didn't answer — see: nsbx status <name>"
 }
 cmd_status(){
-  local name="$1"; mexists "$name" || die "no such sandbox: $name"
+  local name="$1"; mexists "$name" || die "no such sandbox: $name (run: nsbx list to see what exists)"
   python3 -m json.tool "$(manifest "$name")"
-  daemon_alive "$name" && echo "state: running (pid $(daemon_pid "$name")) stats: $(sbx_stats "$name")" || echo "state: stopped"
+  local bpath; bpath="$(sdir "$name")/bin/engram"
+  if [ -f "$bpath" ]; then
+    echo "binary: sha=$(sha "$bpath" | cut -c1-12) built=$(bin_built_at "$bpath")"
+    local fresh; fresh="$(_binary_freshness "$name")"
+    [ -n "$fresh" ] && printf '%s%s%s\n' "$C_YEL" "$fresh" "$C_0"
+  fi
+  case "$(daemon_health "$name")" in
+    running)
+      echo "state: running (pid $(daemon_pid "$name")) stats: $(sbx_stats "$name")";;
+    unresponsive)
+      printf '%sstate: running but NOT RESPONDING%s (pid %s) — process alive, /api/stats returned nothing.\n' "$C_RED" "$C_0" "$(daemon_pid "$name")"
+      info "check: tail -50 $(sdir "$name")/logs/daemon.log   |   next: kill -9 $(daemon_pid "$name") && nsbx up $name"
+      ;;
+    *) echo "state: stopped";;
+  esac
   [ -f "$(sdir "$name")/validate.json" ] && { echo "--- last validation ---"; python3 -m json.tool "$(sdir "$name")/validate.json"; }
 }