Adds tests/native/test_lexer_scaling.el, the regression gate for el #132.
Both directions are proven on LIVE workloads, not synthetic series:
healthy per-character scan 1821 3251 6007 10422 us -> O(n) PASS
rescan-from-zero (the #132 shape) 922 3667 13524 44792 -> O(n^2) FAIL
A gate only proven to pass is decoration. The quadratic specimen exists so
the gate is proven to FIRE.
Also fixes elb_spread_ok to judge the ASYMPTOTIC TAIL (last three ratios)
rather than the whole sweep. Measured on a genuinely linear scan the ratios
ran 3.37 2.92 1.76 1.65 -- the head looks quadratic because it is cold
cache, the tail is the truth. Whole-sweep spread rejected correct data. A
complexity bound is an asymptotic claim and must be judged asymptotically.
That fix came from the classifier refusing to rubber-stamp my own bad
measurement: it reported INDETERMINATE on an unwarmed sweep rather than
passing it. Warmup is now taken and discarded at every sweep point.
Reverts the == workarounds in test_elbench.el now that el #137 has landed;
the natural form generates no str_eq and all 13 fitter tests stay green.
The workaround remains -- the Plus arm is still open.
fn check(label: String, cond: Bool, want: Bool) -> Void {
if cond == want { ... } -> if (str_eq(cond, want)) SIGSEGV
}
Bool has always been an integer in the value model — type_to_c maps Bool to
"int", and el_runtime.h states "Bool -> el_val_t (0 = false, nonzero = true)".
But Bool names were registered NOWHERE: build_int_names_for_params tracked Int
and Float params, and the `let` path tracked Int and Float bindings. Neither
knew about Bool.
So comparing two Bools fell through to str_eq, which dereferenced 0 or 1 as a
char* and segfaulted immediately.
This is the third instance of one family found tonight, after el #137 (a call
on either side of == poisoned the operator) and el #136 (a missing import
compiled clean). All three are the same shape: something the compiler could not
type, silently handled as a string.
Found while writing #137's own test harness — the first version of that harness
crashed on exactly this, on both the old and new compiler, which is how it
surfaced. A test harness that cannot compare two Bools is a good way to notice.
VERIFIED:
- the harness that segfaulted on every prior compiler (exit 139, no output)
now runs clean: 14 passed, 0 failed
- self-hosting fixpoint byte-identical
- the compiler's own generated C differs by 8 lines — only the intended
registration
- neuron's full soul amalgam regenerates in 424ms, exit 0, BYTE-IDENTICAL
- test_math 13/13, test_string 27/27, test_core 10/10, test_text 12/12
Adds tests/runtime/operator_typing_test.el, the 15-case suite from #137, so
this family is covered going forward rather than rediscovered.
let a: Int = 5
getint(5) == a -> str_eq(getint(5), a) SIGSEGV
getint(5) == 5 -> getint(5) == 5 fine
A function call whose return type codegen cannot infer poisoned the operator,
and a declared Int on the other side did not save it. str_eq then read an
integer as a char* and segfaulted. Only an integer LITERAL on one side forced
the numeric form, which is why the bug stayed invisible: the common case
happened to be safe.
The check required BOTH operands to be provably Int:
if is_int_expr(left) { if is_int_expr(right) { numeric } }
Loosening to OR is strictly safer, not a trade:
- when one side is a known Int, str_eq is ALWAYS wrong — it dereferences
that integer — while numeric comparison is at worst a wrong answer on a
program that was already ill-typed;
- when neither side is Int nothing changes at all, so string comparison is
untouched.
Found by the test-framework agent while building the benchmark harness; it
correctly declined to fix it mid-phase since it is a codegen semantics change.
VERIFIED, because a semantics change earns more than an assertion:
- 15/15 on a dedicated operator suite covering string literals, string vars,
string-returning calls, mixed var/call, and != in every combination. The
pre-change compiler scores 0/15 on the same file: it segfaults before
printing anything.
- self-hosting fixpoint byte-identical
- the ONLY difference in the compiler's own generated C is the intended one:
a nested if becoming two sequential ifs, in EqEq and NotEq. Nothing else
moved.
- neuron's full soul amalgam regenerates in 400ms, exit 0, output
BYTE-IDENTICAL at 1,270,212 bytes
- test_math 13/13, test_string 27/27, test_core 10/10, test_text 12/12 —
62 tests, 190 assertions, zero failures
NOT fixed here, same family, flagged for a decision: Bool PARAMETERS are not
tracked as int-like, so `cond == want` between two Bool params still lowers to
str_eq and segfaults. Found while writing this commit's own test harness — the
first version of it crashed on exactly that, on both the old and new compiler.
It needs the same treatment, and it wants its own change.
import "../../NOPE/does_not_exist.el"
compiled CLEANLY — exit 0, empty stderr, and a program silently missing
everything it imported.
resolve_imports did `fs_read(src_path)` and used the result without checking.
fs_read returns "" both for "file is empty" and "file does not exist", so a
typo, a moved file, or a relative path resolved from the wrong working
directory all produced a successful build of nothing.
It caused a real wrong conclusion during test-framework work: a bisection run
from a subdirectory where ../../runtime/ did not resolve produced ELEVEN
consecutive "successful" compiles that had included no runtime at all, and the
results were believed before anyone noticed.
Missing dependency, confident success — the same shape as a test suite
reporting pass for tests that never ran, and as a benchmark reporting 0us
because the optimiser deleted the loop.
fs_exists separates the two cases, so a legitimately empty file still resolves
to "" and is fine. A path that does not exist now prints the resolved path and
exits 1, which is what build scripts check.
Verified:
- bad import: exit 1 (was 0), message names the resolved path
- elc-cli.el still compiles, self-hosting fixpoint byte-identical
- neuron's full soul amalgam regeneration: exit 0, 405ms, output
byte-identical at 1,270,212 bytes
Adds el_black_box (inline asm, +r constraint, memory clobber) and
runtime/elbench.el: a growth-curve classifier that gates time AND
allocation-count AND allocation-bytes, failing if any exceeds its
declared curve.
Refusal is a first-class verdict. The classifier REFUSES rather than
classifying when the largest measurement is below the floor, or when a
series is hard-flat across an 8x input range -- the shape produced when
the optimiser deletes the work. Reporting O(1) there would be a
confident answer with nothing behind it. Disagreeing ratios report
INDETERMINATE rather than a guess.
Deviation from DESIGN.md 6.2, stated in the source: uses consecutive
ratios on a mandated geometric sweep rather than least-squares over
candidate curves. Ratios are directly interpretable on a doubling sweep
and need no floating point; the cost is weaker O(n) vs O(n log n)
separation, reported as an ambiguous band rather than guessed.
Documents the counter scope limit: engram_*.c and libcurl malloc are
NOT tracked, so a flat curve over engram/HTTP-dominated work is not
evidence of anything.
13 tests prove the classifier against real measured series from
fitprobe.el -- including that an accumulator's allocation COUNT is
linear while its bytes are quadratic, and that el #132's pure-CPU shape
reads FLAT on both allocation signals and is caught only by time.
el #131 instrumented the four string allocators, which meant list- and map-heavy
code reported ZERO allocations — a benchmark over lists would have been fitted
against a flat line and passed anything. Caught during framework work: a
"linear" specimen read 0 allocs until it was rewritten to allocate strings.
A gate is only as good as its blind spots are small, and a signal that silently
reads zero is worse than no signal: it produces a confident pass.
Now counted at every container allocation — ElList and ElMap bodies, their
backing arrays, the copy-on-write clones, and the realloc growth path.
Verified on an append loop (n = 100..800):
allocs 7, 8, 9, 10 +1 per doubling = O(log n) reallocations
bytes 2048, 4096, 8192, 16384 exactly 2x per doubling = O(n)
Both curves are what correct amortized growth should look like, and both read
zero before this change.
Known remaining scope, stated rather than left implicit: these counters cover
the runtime's own allocations. They do not see malloc inside engram_*.c or
libcurl, which is correct — the gate is for El-level complexity, not for
third-party memory behaviour.
el #132's quadratic (strlen per character in str_char_code/str_slice) is
pure CPU and allocates NOTHING. Measured on three controlled specimens:
specimen allocs bytes time
linear 2.00 -> O(n) 2.16 -> O(n) 2.05 -> O(n)
accum 2.00 -> O(n) 3.99 -> O(n2) noisy
compute FLAT FLAT 3.96 -> O(n2)
'compute' is #132's shape. A gate fitting only allocation count and bytes
classifies it FLAT and passes -- it would not have caught the defect it
was created for. The gate now fits time AND count AND bytes, failing if
any exceeds its declared curve.
Also: black_box is mandatory and consuming the result is NOT sufficient.
The first 'compute' reported 0us at every n while returning a correct n2 --
clang closed the loop to a multiply. Only an opaque call restored the curve.
Adds lang/tests/bench/fitprobe.el as the fitter's known-good/known-bad set,
so the classifier is provable without depending on a real bug existing.
Marks DESIGN.md 1.3 stale: test_compiler 3.58s -> 0.03s (119x).
el_val_t math_log(el_val_t f) { return el_from_float(log(el_to_float(f))); }
el_val_t math_ln(el_val_t f) { return el_from_float(log(el_to_float(f))); }
Both were natural log, so math_log and math_ln were the same function.
log10(100) returned 4.605 instead of 2.
Three sources already agreed it should be base-10 and were being contradicted
by this one line:
- runtime/math.el:55 "// math_log — base-10 logarithm."
- el_seed.c:1278 __log_f -> log10() (the path math.el actually calls)
- tests/native/test_math.el:133 asserts log10(100) == 2
FOUND BY THE NEW TEST FRAMEWORK ON ITS FIRST RUN (el #133). The assertion had
been sitting in the suite the whole time; nothing could report it. The old
harness printed "N passed, M failed" with no per-test detail, and half the
suites were not compiling at all — so a failing assertion in a suite nobody
could run was indistinguishable from no failure.
That is the entire argument for the framework, demonstrated on day one: this is
not a bug the framework introduced, it is a bug the framework made VISIBLE.
Verified: tests/native/test_math.el goes 12/13 -> 13/13, math-log passing.
Completes the Phase 1 runner and migrates the 11 test files onto it.
- forward-declare the registry accessors in the test preamble; they are
defined at the end of the unit but the El runner is compiled in between
- eltest.el: explicit trailing return in the void emit_* helpers, which
otherwise lower to 'return println(...)' and fail to compile
- test files import runtime/eltest.el explicitly, using the language's own
textual import mechanism rather than compiler-side auto-injection
- DESIGN.md 6.5: gate on allocation COUNT AND BYTES, not count alone
Verified: self-hosting fixpoint byte-identical (gen2 == gen3). 6 of 11
suites run and report per-test timing. The other 5 fail to COMPILE, and
fail identically under the committed compiler -- pre-existing breakage
this framework makes visible for the first time.
THE BUG. str_char_code() and str_slice() each called strlen() on every
invocation. 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).
el_val_t str_char_code(el_val_t s, el_val_t i) {
...
int64_t n = (int64_t)strlen(str); // <- O(n), every call
if (idx < 0 || idx >= n) return 0;
return str[idx];
}
HOW IT WAS FOUND. Not by reading code — by sampling the running process, which
is the same method that resolved tonight's engram outage after four wrong
theories. A geometric sweep of synthetic sources showed wall-clock rising 3.0x,
3.0x, 4.0x, 4.14x per doubling (converging on 4x = quadratic), and a stack
sample put 779 of 779 samples inside lex(), every one bottoming out in
_platform_strlen via str_char_code and str_slice.
THE FIX. Remember the length instead of recomputing it. The subtlety is
INVALIDATION: El strings are arena-allocated, so a freed pointer can be reused
for a different string at the same address, and a naive pointer-keyed cache
would hand back a stale length and read past the end of the new string —
trading a performance bug for a memory-safety one. So entries carry a
generation, a hit requires pointer AND generation to match, and every path that
frees or mutates a runtime string bumps the generation: el_arena_pop,
seed_request_end, __str_set_char. Stale entries cannot be believed; they miss
and recompute.
MEASURED, same host, same inputs:
n(fns) before after
512 0.10s 0.01s
1024 0.37s 0.02s
2048 1.51s 0.03s 50x
the compiler's own 422 KB source concatenated (DESIGN.md's 3.58s case):
3.55s -> 0.03s 118x
The speedup GROWS with input size, which is the signature of removing a
complexity class rather than a constant factor. After the fix each doubling
adds ~0.01s: linear.
CORRECTNESS, verified rather than assumed:
- byte-identical output on every sweep input (n = 128..2048)
- byte-identical output on the 422 KB compiler concatenation
- byte-identical output on tests/runtime/string_test.el
- self-hosting fixpoint byte-identical
- new tests/runtime/str_cache_test.el: 17 assertions covering bounds, empty
strings, negative indices, slice clamping, distinct strings not sharing a
cached length, 1000 interleaved strings forcing cache-slot collisions, and
a grown string not reporting its old length. All pass.
This is the defect that made dist/soul.c a committed artifact: elc could not run
in CI because it needed 24 GB+ and minutes. It needs neither now.
Replace the hardcoded test harness main() with a generated static registry
and index-based accessors, and move all reporting into runtime/eltest.el.
The old harness inlined direct calls into main() and counted assertions in
two globals. That shape cannot report which test failed, how long any test
took, or whether a test ran at all -- a misspelled registration reported
success for a test that never executed.
- assertions record into per-test state instead of global counters
- registry table emitted at compile time; discovery strictly precedes
execution, which is what later enables --list, filtering and sharding
- per-test wall timing on CLOCK_MONOTONIC, taken in C around the call
- runner in El: structured NDJSON events as source of truth, human output
rendered from the same fields
Implements the three primitives the test-framework design (DESIGN.md §6.5)
requires for gating on growth curves: el_alloc_count, el_alloc_bytes,
el_peak_rss. Registered in codegen's builtin_arity and wrapped in el_seed.c per
the project's C-builtin recipe.
WHY COUNTS AND NOT WALL-CLOCK: a growth-curve gate has to be a hard build
failure, which means the signal cannot flake. Wall-clock needs warmup,
statistics, and a quiet machine; on shared CI it is unusable as a gate.
Allocation counts are perfectly deterministic — same input, same number, every
machine, every run. Fit them against n and a complexity regression becomes a
build failure with zero noise.
All four runtime string allocators (el_strdup, el_strbuf, and their _persist
variants) funnel every allocation the language performs, so instrumenting there
counts everything.
WHY BYTES AS WELL AS COUNT — this is not redundancy, it is the whole gate.
Measured with two El programs, one allocating once per item, 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 — identical to the
healthy one. Counting allocations alone would have missed it completely. Bytes
catch it: each doubling of n quadruples bytes (ratios 3.94, 3.97, 3.99 ->
converging on 4.0, i.e. O(n^2)), while the linear case converges on 2.0.
That shape — count linear, per-allocation size growing — is the classic
accidental quadratic, and it is exactly elc's defect: quadratic allocation
VOLUME, which the old shipped compiler paid in RSS (27 GB, OOM) and the rebuilt
one pays in malloc/free churn (42s on 1.4 MB). Volume was the invariant across
both; RSS and wall-clock were just the two ways it surfaced.
el_peak_rss is exported for context and is explicitly NOT a gating signal — it
is perturbed by allocator internals, the page cache, and the OS. Gate on the
deterministic numbers; report the physical one.
Counters are unsynchronised by design: this is measurement, and a lock would
change the thing being measured. Exact on the single-threaded compile path,
approximate under threads.
The guard I added minutes ago checked swap availability as a level
(avail < total/8 -> report zero available). That is the wrong signal, and the
same host proved it twice within minutes:
47.65 / 48.00 GiB swap used, 2047 swapouts/s -> genuinely thrashing
26.67 / 28.00 GiB swap used, 0 swapouts/s -> healthy, 15.6 GiB free
Both are ~97% "used". macOS grows swap files on demand and trims them lazily,
so the level says almost nothing about now — it is a high-water mark. The level
check calls the second state an emergency and starves the pool for no reason,
which is its own failure mode: a guard that fires on healthy machines gets
disabled, and then guards nothing.
What separates the two is whether pages are moving. So sample the swapout
counter across calls and judge the delta:
- > 200 pages/s (~3 MiB/s) sustained outward paging => report zero available;
callers refuse to grow and pc_relieve_pressure hands frames back.
- The first call primes the baseline and reports no pressure. One sample
cannot have a rate, and inferring one from a single reading is exactly the
mistake this commit removes.
Measured thresholds, not guessed: idle sat at 0/s, recovery burst hit 24,845/s
while the compressor drained (transient, correctly not a growth decision since
growth is only evaluated on eviction passes), and real thrash held ~2000/s.
200/s sits clearly above noise and far below either.
The compressor-footprint subtraction stays: that RAM is genuinely spoken for
regardless of paging rate.
The adaptive budget I added an hour ago could only grow, and grew toward a
share of TOTAL ram (80%, ~38 GiB on a 48 GB host). That is a memory leak with
extra steps: total never shrinks when other processes need memory, so the pool
had no way to notice it was starving the machine it runs on. Deployed briefly;
caught as memory pressure on the host.
A control loop with only one direction is not a control loop.
- pc_available_ram(): free + inactive + purgeable via host_statistics64 on
Darwin, MemAvailable on Linux. Availability is the quantity that moves when
the machine is under pressure; total is not. Returns 0 when it cannot be
read, and callers then refuse to grow — a cache is never worth swapping the
host, so unknown means no.
- Growth is bounded by availability minus a free-memory floor (2 GiB default,
ENGRAM_POOL_FREE_FLOOR_MB), not by total. The share-of-total ceiling stays
as a second bound and drops 80% -> 50%.
- pc_relieve_pressure(): the missing direction. On every eviction pass, if
available memory is under the floor, hand back ~25% of held frames; the
resident set follows on the next pass so the memory is actually returned
rather than merely re-labelled. Counted as adapt_shrinks alongside
adapt_grows so both directions are visible in the same report.
- pc_default_cap() also clamps the STARTING budget to what is spare right
now, so a cold boot on a loaded machine does not open at a size the host
cannot afford.
Verified on a 48 GB host: engram boots in ~30s, RSS settles at 2.22 GiB (the
store's actual size, resident, not creeping), 0.0% CPU, 13,439 nodes / 37,670
edges, embeddings complete. Guard reports 9.71 GiB available against a 2.00 GiB
floor — 7.71 GiB of headroom it is permitted to use and no more.
Follow-on to the edge write barrier. That fix removed the full-store walk;
this one makes the pool able to notice if anything like it happens again.
WHAT WENT WRONG, precisely: the pool thrashed the live engram to a standstill
twice on 2026-08-15 and said nothing. From outside it was indistinguishable
from "busy loading" — 100% CPU, flat RSS, no output — so four wrong theories
got tried (bad binary, corrupt snapshot, WAL replay, feature flags), each
costing a deploy or a rollback. The whole time, hits/misses/evictions were
already being counted in PgCache, and the struct comment read:
/* stats (introspection only — never affect semantics) */
That comment was the bug. Self-measurement treated as decoration is why the
pool could not correct itself and why no one outside could see what it was
doing. A system that cannot read its own state cannot correct, and neither can
anyone watching it.
- pc_adapt_budget(): the loop, closed. Over a sliding window, evictions
running at a large fraction of accesses WHILE reuse is real means the
working set exceeds the budget — so grow it, geometrically, bounded by a
LIVE re-read of physical memory. Evictions alone are not pressure (a scan
evicts and never returns); evictions with reuse are. An explicit
ENGRAM_POOL_FRAMES still wins — an operator override must not be silently
overruled.
- Budget derived, not declared. A constant cannot be right: 16 GiB of frames
is arbitrary on a 48 GB host and suicidal on a 16 GB one. Even "60% of RAM
at startup" is a guess about the future — it cannot know the store grew or
the machine changed. Hence the live re-read.
- pc_report(): ONE structured emission carrying the entire sensed state,
through emit_log — El's existing telemetry, already exporting to OTLP.
Deliberately not a function per stat, and deliberately not a bespoke
/api/pool endpoint: both make observability something hand-written per noun
instead of the uniform mechanism every component already has.
- engram_pool_stats_json(): the same state readable live, wired through the
normal builtin path (codegen arity + el_seed wrapper), so the pool can be
observed in real time rather than reconstructed afterward from a stack
sample.
Verified: with the exact configuration that took production down
(ENGRAM_POOL_FRAMES=65536 → 1 GiB cache against a 2 GiB store) the engram boots
clean and serves — 0.0% CPU, 13,436 nodes / 37,663 edges, embeddings complete —
and NO pressure event fires, because the barrier removed the walk that caused
it. The controller is defense in depth; the barrier is the fix.
Checkpointing pushes the ENTIRE resident graph through store_put_node and
store_put_edge (see engram_store_checkpoint). Nodes were cheap: a durable-hash
compare skipped unchanged records with zero page I/O. Edges had no barrier at
all — struct comment at PgCache.barrier_on even says "node durable-hash
barrier" — so every edge was rewritten on every checkpoint, and each rewrite
runs the idempotency probe max_page_lsn_for_id -> btree lookup -> page_read.
Edges outnumber nodes ~3:1 here (37,663 vs 13,436), so routine checkpointing
degenerated into a FULL-STORE WALK in id order: random page access across the
whole 2 GiB store, repeated, overwhelmingly to rediscover nothing had changed.
LRU is worst-case under exactly that pattern — it evicts the page it is about
to want — so once the page cache was smaller than the store, the walk collapsed
into thrashing: 100% CPU, flat RSS, no forward progress, port never bound.
That took the live engram down twice on 2026-08-15.
The walk is the defect. Sizing the cache to survive it treats the symptom.
Changes:
- dh_edge_hash(): edge counterpart of dh_node_hash, with a kind discriminator
byte so an edge can never collide with a node of the same id in the shared
map. created_at/updated_at/last_fired are excluded deliberately: last_fired
is touched by activation without changing what the edge IS, and folding it
in would defeat the barrier on precisely the hot edges that most need it.
- store_put_edge(): barrier check + dh_set on success, mirroring
store_put_node exactly.
- store_scan_edges(): seed the barrier map from on-disk truth at load, so the
FIRST post-boot checkpoint already skips unchanged edges. store_scan_nodes
already did this and its comment says why; edges were simply never done.
Verified: with the exact configuration that killed production
(ENGRAM_POOL_FRAMES=65536 -> 1 GiB cache against a 2 GiB store), the engram now
boots clean and serves — LISTENING, 13,436 nodes / 37,663 edges, embeddings
complete, 0.0% CPU, RSS 1.14 GiB (cache resting at its budget rather than
thrashing against it). Same small cache, same store, no walk.
/api/graph/edges answered a read query by calling engram_save() to serialize
the ENTIRE graph to disk (128 MB) and then fs_read-ing it back. Two defects in
one line, and both bit production on 2026-08-15:
1. The path it wrote was ~/.neuron/engram/snapshot.json — the engram
server's CANONICAL store. A READ route overwriting the persistence
owner's canonical file. This defect had been fixed once (export moved to
a scratch path); it came back when the hand-written dispatch block was
replaced by @route dispatch and the unfixed copy is the one that
survived the merge.
2. Cost: a full snapshot write, a 128 MB read, and a parse of the whole
graph, per request, to return a bounded slice.
Calling it tonight overwrote the canonical snapshot and immediately preceded
an engram crash loop.
engram_edges_json(limit, offset) is the builtin that route's own TODO asked
for ("Future: add an engram_edges_json() builtin and drop the file round trip
entirely"). It walks g->edges directly and emits every persisted field.
limit <= 0 defaults to 1000, not unbounded: this is the endpoint that fell
over, and an unbounded default would preserve the failure mode under a new
name. Callers page explicitly.
Registered in codegen.el's builtin_arity (both plain and __ spellings) and
wrapped in el_seed.c per the project's C-builtin recipe.
neuron's soul calls engram_recall_json (neuron-api.el:618, memory.el:80) and
cgi_principal (studio.el:72). Both existed in the runtime neuron vendored
(v1.0.0-20260501) and were absent here, so the soul could not link against
current el at all.
The dangerous part is what the obvious "fix" would have done. These look like
redundant wrappers over one impl:
engram_search_json(q, limit) -> eg_search_json_impl(q, limit, 0) LEXICAL
engram_recall_json(q, limit) -> eg_search_json_impl(q, limit, 1) SEMANTIC
They are not interchangeable, and the split is documented at neuron-api.el:613:
search stays LEXICAL because ~40 internal call sites pass a KEY and seven of
them DELETE every record returned. Point those at a semantic matcher and they
delete fuzzy matches. Conversely, pointing recall at search silently downgrades
the mind's entire retrieval surface from semantic to lexical — no error, just
permanently worse recall.
Implemented over engram_activate(), which in this runtime already IS the
semantic path the old with_legs=1 branch built by hand (embeds the query via
eg_embed_fetch, scores by cosine, then spreads activation one hop). Output
shape matches engram_search_json — a flat array via engram_emit_node_json —
because callers parse search's shape, not activate's envelope.
Verified: neuron's soul now compiles and links against current el, boots, and
serves /health with layers initialized.
NOTE for follow-up: current el also ships engram_retrieve_geometric_json, a
structure-first retrieval that appears to be the intended successor to recall.
Repointing the two recall call sites at it may well be the right end state and
would remove the two-wrapper shape entirely — but that is a behavioral change
that must be measured against neuron/tools/retrieval-eval/'s gold set, not
assumed. This commit preserves existing behavior exactly; it does not decide
that question.
The committed elc binary could not be refreshed from its own source. Rebuilding
failed with three implicit-declaration errors: el_mem_check, stdout_to_file,
stdout_restore. The compiler's own source calls all three (compiler.el:472,479,574
and codegen.el:4248) and two are registered in codegen.el's builtin_arity table —
but none were defined in this runtime.
They were found intact in ui/examples/native-hello-ios/NativeHello/el_runtime.c,
a divergent private copy of this runtime that still carried them. Ported verbatim.
Consequence of them being missing: the canonical elc binary was frozen. Source
gained @route dispatch codegen (emit_route_dispatch, codegen.el:3948) and the
@manager boundary-beat seam, but no rebuilt binary could carry them, so
neuron's soul — whose routes.el now calls the compiler-synthesized
el_route_dispatch — could not be built at all.
Verified after the fix:
- elc rebuilds from current source, clean.
- Self-hosting fixpoint byte-identical (stage3 == stage2).
- The rebuilt elc emits el_route_dispatch (2 occurrences in the soul amalgam,
previously 0) and injects engram_boundary_beat at @manager boundaries,
i.e. the decorator seam is live rather than inert.
el_mem_check is itself the compiler's memory guard (ELC_MAX_MEM_MB, default
512MB, self-terminates before the OS OOM-killer fires) — so the runtime was
missing the very guard that would have surfaced the compiler's memory blowup
as a clean error instead of a 27GB host-killer.
Three real bugs, all found by actually running the thing rather than reading it.
1. query_param never URL-decoded. A GET of /api/search?q=neural%20network
searched for the literal string "neural%20network" and returned []. Every
multi-word search against the live engram has been silently returning empty
results — not an error, an empty result, which is why it went unnoticed.
Affects every GET route that reads query params, not just search.
2. query_param matched key names unanchored. str_index_of(qs, "q=") matches
inside "faq=", so "?faq=X&q=Y" returned X for key "q". Verified live before
the fix. Now searches for "&key=" against "&"+querystring so a match can
only land on a real parameter boundary.
3. el_request_start/el_request_end were defined in BOTH el_seed.c and
el_runtime.c, so linking the two objects together — which is exactly what
the product build does — failed with duplicate symbols. el_seed.c's own
comment already says these moved there ("formerly defined in el_runtime.c.
Now self-contained in el_seed.c"); the el_runtime.c copies were left behind
during that move. Removed them, kept declarations since http_worker calls
them. Also added the three missing prototypes (engram_op_assert_json,
engram_node_full_in, engram_connect_in) that el_seed.c wraps but never
declared, which made it fail to compile standalone under C99+.
Verified: engram builds and links clean from canonical source; before/after
comparison on a copy of the real store shows "neural network" returning a real
match where the live build returns [], and "?faq=WRONG&q=MetaColloc" now
resolving to MetaColloc. Live engram on :8742 was never touched.
#116 shipped the ggml strategy at 0.9933 id-recall against the CPU oracle
while the hand-rolled Metal kernel it replaced scored 0.9997 — a ~150x worse
error margin. That was not an inherent property of ggml. It was a usage bug in
this file, and this commit fixes it.
ggml-metal has two F32xF32 matmul kernels and picks between them purely on
ne11, the number of B rows, which for us is the query-batch size:
ne11 <= 8 -> kernel_mul_mv_ext_f32_f32_* / kernel_mul_mv_f32_f32_*,
templated <float, float> — genuine F32.
ne11 > 8 -> kernel_mul_mm_f32_f32, templated
<half, half4x4, simdgroup_half8x8, half, half2x4, ...> —
BOTH operands narrowed to F16, despite F32 tensors on both
sides.
The old code issued one ggml_mul_mat with ne11 = nq (300 in the benchmark),
landing squarely on the F16 path. The file's own header comment asserted the
opposite ("computes in F32 on the Metal backend"); that claim was wrong and is
replaced with the measurement.
Fix: emit ceil(nq/8) mul_mats over ne11<=8 ggml_view_2d slices of one query
tensor, all expanded into ONE graph and one ggml_backend_graph_compute, so the
node matrix is still uploaded and shared exactly once. EL_GGML_MULMAT_CHUNK
overrides the 8; setting it >= nq reproduces the old behaviour exactly, which
is also how the before/after below was measured in a single binary.
Measured, real store snapshot, 13415 live embedded nodes, dim=768, 300 real
queries, vs the CPU double-accumulated oracle (vindex_bench, offline copy of
the store — no live service touched):
id-recall same-rank |Δdist| max mean
old (ne11=300) 0.9933 6.80e-05 1.43e-05
new (ne11<=8) 0.9987 4.77e-07 9.30e-08
hand-rolled 0.9997 3.58e-07 7.55e-08
~145x better max error, ~154x better mean — now the same order of magnitude as
the hand-rolled kernel rather than 150x off it.
The cost is real and is documented rather than buried. Median of 15 reps of
the whole batch_multi() call, three runs: 13.2-14.4ms unchunked, 19.9-20.2ms
chunked, 17.7-18.0ms hand-rolled. Correctness costs ~+6.7ms per 300-query
batch and leaves ggml ~12% behind the hand-rolled kernel instead of ~35%
ahead. It cannot be recovered inside ggml: an fp32 matmul on Metal must
re-stream the node matrix once per <=8 queries, and ggml's Metal backend ships
no fp32 TILED matmul, so "fast" and "fp32" are genuinely exclusive there.
Two things that did NOT work, recorded so nobody retries them:
- ggml_mul_mat_set_prec(t, GGML_PREC_F32) does nothing here. Error was
bit-identical with and without it (1.038e-05 either way) — ggml-metal has
no F32-accumulating mul_mm kernel to switch to. ne11 is the only lever.
- The ACCEL/BLAS device looked excellent in an isolated compute-only probe
(3.4-4.0ms, mean |Δdot| 1.5e-08) but is dominated on BOTH axes end-to-end
(0.191 ms/query at 0.9973 recall vs 0.125-0.142 at 0.9987), because the
probe was not competing for the same CPU cores the real call path is. It
stays reachable via EL_GGML_DEVICE as a no-Metal fallback, labelled as
measured-and-rejected, not as a recommendation.
Also corrected: the ~7.8s "cold start" blamed on this file is not this file
re-initialising per call — init was already cached. It is Apple's shader cache
missing on ggml's embedded metallib (~650 kernels), keyed on the library and
shared across processes: the first load on a machine reports
"loaded in 7.670 sec", the next run of a *different* binary reports 0.009 sec.
Once per machine per ggml version, not once per process, and not ours to fix.
Warm ggml init is 44-53ms vs 36-117ms for the hand-rolled strategy.
Loading only libggml-metal.so instead of every plugin in the directory is kept
for tidiness, and explicitly documented as NOT a speedup: 44.7-52.4ms against
46.9-58.9ms, the same number inside noise.
The -2.0 sentinel contract is unchanged and re-verified at batch sizes that
straddle the chunk boundary (1,7,8,9,16,17,33), plus NULL rows, dim
mismatches, zero-norm rows, and an all-invalid population. Notably the old
ne11=300 path fails that same check at a 2e-6 cosine tolerance with 2299
mismatches, which is an independent confirmation of the defect.
runtime/channel.el has always called __channel_new/__channel_send/
__channel_recv/__channel_try_recv/__channel_close, but these were only ever
implemented in the pre-restructure lang/el-compiler/runtime/el_runtime.c.
When the canonical runtime was consolidated onto the release copy
(lang/runtime/el_runtime.c) and el_seed.c became the sole C dependency,
the channel implementation was never carried forward — __mutex_new made the
move, __channel_* did not. Any El program using Go-style channels currently
fails to link on dev.
Ported the working buffered-MPMC-channel implementation (mutex+condvar+
circular buffer, bounded and unbounded modes) from the old el_runtime.c
verbatim, adapted only to el_seed.c's arena API (seed_arena_track in place
of el_arena_track). Declared in el_seed.h alongside the existing mutex
primitives.
transduce() is now THE single mechanism: one function, no content-type
branch inside it. It never asks whether `source` is prose, JSON, or
raw/opaque bytes (audio, etc.) — it runs one algorithm unconditionally:
split on "\n\n" as a universal boundary-marker check, and if that finds
no boundary, fall back to fixed 4096-char windows. Same node/edge wiring
(root -contains-> chunk, chunk -precedes-> next, "#"-prefixed chunk gets
a heading/section_of link) regardless of what's inside a chunk. Dedup is
the existing find_existing_by_content path via merge_manifold, applied
uniformly. The old transduce_structured JSON dataset/records/feature-node
interpretation is deleted outright, not just unused — a JSON file now
gets chunked and deduped like anything else, with no pre-computed
structure. All five ingest_* entry points still exist unchanged in name
and role; ingest_file/ingest_dir/ingest_url/ingest_llm now call the one
transduce() (ingest_stream builds its own turn-nodes directly and never
called either old function, so it's untouched).
This unlocks raw/opaque content (audio, or anything else with no natural
text/JSON shape) without any DSP, LLM call, or external API: transduce()
chunks it exactly like it chunks anything else. There is zero semantic
understanding of audio (or any payload) claimed or built here — any
meaning is expected to emerge later from Neuron's own existing mechanisms
(embedding, spreading activation, dedup) acting on this real geometry
over time.
Two small C builtins added to el_runtime.c/h (fs_size, fs_read_b64_chunk)
because El strings are NUL-unsafe under strlen-based ops and fs_read()'s
result silently truncates at the first embedded NUL, which is routine in
real binary/audio bytes. ingest_file compares fs_read()'s string length
against a real fs_size() stat() count; on mismatch it rebuilds the
payload as base64-encoded fixed 3072-byte windows read directly off disk
(binary-safe in C, verbatim, no invention), joined with the same "\n\n"
marker transduce()'s boundary scan already looks for. This is a
mechanical fidelity fix, not interpretation of content — transduce()
never learns a fallback happened. Registered both builtins' arity in
codegen.el; did not rebuild the elc compiler binary itself (unrelated,
pre-existing gap: self-hosting elc via el_seed.c fails on this worktree
independent of this change, reproduced with codegen.el reverted) — the
existing elc binary compiles calls to unregistered builtins via its
already-existing arity=-1 passthrough, confirmed by an actual clean
`elc ingest.el` + `cc` build against the modified el_runtime.c.
INGEST_KIND keeps existing only as an acquisition-mechanism selector
(dir/file/url/llm/stream — which RPC to use to fetch bytes), not as a
content-type flag; the redundant "structured" value (an alias for "file"
that hinted the now-deleted JSON branch) is removed. ingest_dir drops its
file-extension filter for the same reason: transduce() takes anything now.
Verification: local manifold construction confirmed correct against a
real captured audio file (will_clean.wav, 304288 bytes, and a 12288-byte
real prefix slice) — exact expected node/edge counts both times
(101 nodes/199 edges full file; 5 nodes/7 edges for the slice, matching
ceil(bytes/3072)+1 nodes and 2n-1 edges), with real, verbatim base64
content confirmed decoding back to the actual WAV header bytes. Compiles
clean via the real elc + the modified el_runtime.c/engram_*.c (built and
booted an actual sandbox engram off this exact source with `nsbx create
--branch`).
NOT verified this session, disclosed rather than papered over: end-to-end
server-confirmed persistence (a real before/after /api/stats delta, and a
fetched node by id) for the audio, prose, and JSON-fixture cases. Every
local nsbx sandbox engram tried tonight (two stock pre-#109 binaries
hitting the known O(N*D) brute-force scan bug, then a fresh #109/HNSW
binary built from current dev) took minutes-to indefinitely long on the
final /api/load-merge write's embedding step and hit the client's 60s
HTTP timeout before responding, even for a 5-node write. This is
confirmed as real (if slow) forward progress, not a hang: the sandbox's
WAL file was observed growing steadily across every attempt. The code's
own pre-existing HONESTY GATE correctly refused to report success in
every case, returning "load-merge failed: ..." with a
"nothing below this manifold was confirmed persisted by the server" note
instead — exactly as designed. This is an environment/infrastructure
limitation, not a defect introduced by this change: the engram server
binary itself is untouched by this commit.
will.anderson
Neuron