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.
Reconciles PR #105 ("fix: engram search latency — pin embed model, cache
query embeddings, bound activate BFS") with dev's ACTUAL current
engram_activate, rather than the ancient pre-restructure snapshot #105 was
built against.
WHY THIS NEEDED RECONCILIATION, NOT A DIRECT PORT: #105's single commit
(1dc49b1) modifies `lang/el-compiler/runtime/el_runtime.c` — a path that does
not exist on dev (dev has `lang/runtime/el_runtime.c`; the restructure that
renamed it happened after #105's branch point, which traces to a July 22
merge-base, weeks before the M8/M8.1/qgate/fan-effect/adjacency-index work
this file has grown since). #105's own engram_activate is consequently the
PRE-restructure version: no adjacency index (O(E) full edge scan per hop),
no query-aware qgate, no ACT-R fan effect, no eg_edge_eff_weight, and no
awareness of dev's cosq/e_eff embedding-blend semantic layer — it built a
parallel `g_qcache`/`engram_embed_raw` mechanism from scratch against code
that no longer exists at that path. A raw merge/cherry-pick was not possible
and would have been wrong even if it were: taking #105's tree wholesale would
have thrown away everything dev grew in the meantime (qgate, fan effect,
adjacency index, and this session's own M8 HNSW vindex integration).
RECONCILIATION: kept dev's cosq/e_eff mechanism as the semantic layer
entirely intact (unchanged by this commit) and ported #105's three genuinely
additive wins on TOP of it, at their equivalent sites in the CURRENT
eg_embed_fetch/engram_activate:
1. keep_alive:-1 on the Ollama embed request body (eg_embed_fetch) — pins
the embed model resident so a larger generation model loading under
unified-memory pressure can't evict it and force a cold reload on the
next search (#105 measured ~2.2s cold vs ~0.02-0.05s warm).
2. Query-embedding cache upgraded from dev's single-slot (`_eg_qcache_text`,
only ever remembered the LAST query) to a direct-mapped, FNV-1a-keyed,
1024-slot cache (reusing the existing engram_id_hash) — so the
curiosity loop's rotating phrases actually hit the cache instead of
evicting each other every call. Same "pointer owned by the cache, not
freed by caller" contract as before, just per-slot instead of global.
3. Beam cap on the layer-1 spreading-activation BFS (new
engram_activate_beam(), tunable via ENGRAM_ACTIVATE_BEAM, default 128).
The FIFO frontier is processed in hop-level batches (entries sharing
.hops are provably contiguous — see the code comment); when a level
exceeds the beam width, only the top-`beam` by activation actually
EXPAND. Every node in an oversized level still gets reached[]/best_bg[]
recorded (that happens at enqueue time, one level up) and appears in
the reported/promoted set — the cap bounds associative SPREAD width
only, never recall of what was already found. Kept as a genuine
additional bound even though the adjacency index + qgate + fan effect
already mitigate #105's original "hub-node explosion" failure mode for
a different reason: those prune WHICH targets matter; this bounds
worst-case width regardless.
Everything else in dev's engram_activate — cosq/e_eff, the qgate rescale,
the fan effect, eg_edge_eff_weight, the M8 HNSW vindex seed discovery from
the #109 reconciliation earlier this session — is untouched.
VERIFIED (nsbx sandbox only, live :8742/:7770 never touched): cc -std=c11
-O2 clean build; booted in an isolated sandbox against a real cloned
production snapshot (13,424 nodes / 37,656 edges); ran 5 activate() calls
across rotating queries at depth 3, including the same query issued twice
non-consecutively (2nd hit landed at 476ms vs the 1st at 483ms — consistent
with a cache hit once Ollama's own warm-model latency is accounted for; no
crash, correct varied result counts (367-2610 nodes) each call; act-stats
JSON read correctly throughout.
Built on top of the M8/#109 reconciliation (bacaf3d, merged to dev as
#109) — dev's current HEAD at the time of this commit.
runtime/el_seed.c does not compile standalone via the exact command
tools/install.sh uses (`cc -std=c11 -O2 -I runtime -c runtime/el_seed.c`):
51 of its __-prefixed wrapper functions (http serving, JSON access, key-val
state, URL/HTML escaping, and the whole engram_* node/edge/layer/search
surface) call unprefixed counterparts that are implemented in el_runtime.c,
not in el_seed.c itself, and el_seed.c never declared them -- a toolchain
that treats an implicit function declaration as a hard error under C11
fails the compile outright.
install.sh already 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.
A plain `#include "el_runtime.h"` was tried first and rejected: it redefines
el_to_float/el_from_float, which el_seed.h already provides -- a real
compile error, not a style preference. Added narrow prototypes instead,
copied verbatim from el_runtime.h, for exactly the 51 symbols el_seed.c's
wrappers reference and nothing else.
Verified clean:
- `cc -std=c11 -O2 -I runtime -c runtime/el_seed.c` (install.sh's exact
per-file compile) -- 0 errors, 0 warnings, even with -ferror-limit=0.
- full `tools/install.sh` run -- compiles both objects and archives them
into libel.a successfully.
Separately (not fixed here, out of scope): AGENTS.md's documented compiler
self-rebuild command links elc-new.c against el_seed.c, but elc-new.c's own
generated `#include "el_runtime.h"` line and 3 undeclared symbols
(el_mem_check, stdout_to_file, stdout_restore -- present in neither
el_runtime.c nor el_seed.c) mean that command fails regardless of which
runtime file it's linked against; and install.sh's libel.a only archives
el_seed.o + el_runtime.o, so any program that calls into the engram_*
surface fails to link against it (el_runtime.c's engram_* wrappers need
engram_store.c/engram_geometry.c/engram_reason.c/engram_cognition.c/
engram_vindex.c, none of which install.sh compiles in). Both are real,
pre-existing, and independent of this fix -- worth their own look.
Three confirmed-live bugs tonight:
- `nsbx up` printed "daemon did not become ready" immediately followed by a
green "your sandbox is ready" banner and exited 0, because the existing-
sandbox restart path (`daemon_alive || start_daemon`) never checked
start_daemon's return code. `cmd_build` had the identical unguarded
pattern, plus `cmd_run`/`cmd_validate`'s own start-if-dead calls. All four
now `|| die` with a message pointing at daemon.log.
- `nsbx status`/`nsbx list` reported bare "state: running" for a process
that's alive (passes kill -0) but not actually answering /api/stats --
pegged, hung, or mid-boot. Added daemon_health(), which does the real
stats fetch and distinguishes stopped/running/unresponsive; both commands
now say "running but NOT RESPONDING" with a next-step hint instead of
silently going quiet on the stats field. Reproduced live against another
agent's actively-running (CPU-pinned, non-responsive) sandbox tonight, and
again via a deliberate SIGSTOP on a throwaway sandbox.
- Sandboxes carried no visible signal that their binary predated a relevant
fix. `status`/`list` now show the binary's sha + real build timestamp
(mtime survives `cp -p`), plus a best-effort staleness note: for
stock-prod clones, compare against the currently-configured live binary;
for source/branch builds, compare the recorded source commit against
local origin/dev via merge-base --is-ancestor.
Also, found live while verifying the above:
- A cold boot under concurrent sandbox/CPU load can legitimately take past
the old hardcoded 15s readiness window. Made it configurable
(NSBX_READY_TIMEOUT_SECS) rather than just widening the default blindly.
- cmd_create's post-boot baseline capture could silently record sbx_baseline
as 0/0 when the stats fetch came back empty right after the auto-remerge
step -- which would make every future `nsbx validate` zero-loss/reboot-
prove check trivially PASS regardless of real data loss. Added a bounded
retry and a loud warning if it still comes back empty.
- Sharpened a handful of "no such sandbox" / missing-binary errors to name
the next command instead of just stating the failure.
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.
Stop hand-rolling GPU kernels for batch cosine similarity — use ggml (the
MIT-licensed compute library underneath llama.cpp, installed standalone via
Homebrew) as the preferred backend, without ripping out PR #114's
carefully-verified hand-rolled Metal shader.
Structure: one stable public adapter (eg_cosine_batch.h, zero #ifdef at call
sites) backed by three selectable concrete Strategies behind an internal
vtable (eg_cosine_batch_strategy.h) chosen by a Factory (eg_cosine_batch.c):
- eg_cosine_batch_strategy_ggml.c — NEW. ggml + dynamically-loaded Metal
backend plugin (ggml_backend_load_all_from_path
+ ggml_mul_mat for the batched dot
product), gather/scatter around the
-2.0 sentinel contract.
- eg_cosine_batch_strategy_metal_hand.m — PR #114's original hand-rolled
Metal shader bridge, preserved
almost verbatim, now one strategy
among several rather than the only
option. eg_cosine_batch.metal kept
byte-identical to the original.
- eg_cosine_batch_strategy_cpu.c — universal always-false fallback
(direct descendant of PR #114's
eg_metal_cosine_stub.c).
Selection: EL_COSINE_BATCH_STRATEGY=ggml|metal|cpu|auto (default: ggml first,
then hand-rolled Metal, then CPU — first available wins), plus back-compat
EL_METAL_COSINE=0 to disable every GPU-backed strategy. build_vindex_bench.sh
compiles all three strategies on Darwin, CPU-fallback-only elsewhere.
vindex_bench.c now reports BRUTE-GGML and BRUTE-METAL side by side against
the same CPU oracle, on the same dataset, in one run (real numbers vs. real
store snapshot in the PR body).
Reconciles PR #105 ("fix: engram search latency — pin embed model, cache
query embeddings, bound activate BFS") with dev's ACTUAL current
engram_activate, rather than the ancient pre-restructure snapshot #105 was
built against.
WHY THIS NEEDED RECONCILIATION, NOT A DIRECT PORT: #105's single commit
(1dc49b1) modifies `lang/el-compiler/runtime/el_runtime.c` — a path that does
not exist on dev (dev has `lang/runtime/el_runtime.c`; the restructure that
renamed it happened after #105's branch point, which traces to a July 22
merge-base, weeks before the M8/M8.1/qgate/fan-effect/adjacency-index work
this file has grown since). #105's own engram_activate is consequently the
PRE-restructure version: no adjacency index (O(E) full edge scan per hop),
no query-aware qgate, no ACT-R fan effect, no eg_edge_eff_weight, and no
awareness of dev's cosq/e_eff embedding-blend semantic layer — it built a
parallel `g_qcache`/`engram_embed_raw` mechanism from scratch against code
that no longer exists at that path. A raw merge/cherry-pick was not possible
and would have been wrong even if it were: taking #105's tree wholesale would
have thrown away everything dev grew in the meantime (qgate, fan effect,
adjacency index, and this session's own M8 HNSW vindex integration).
RECONCILIATION: kept dev's cosq/e_eff mechanism as the semantic layer
entirely intact (unchanged by this commit) and ported #105's three genuinely
additive wins on TOP of it, at their equivalent sites in the CURRENT
eg_embed_fetch/engram_activate:
1. keep_alive:-1 on the Ollama embed request body (eg_embed_fetch) — pins
the embed model resident so a larger generation model loading under
unified-memory pressure can't evict it and force a cold reload on the
next search (#105 measured ~2.2s cold vs ~0.02-0.05s warm).
2. Query-embedding cache upgraded from dev's single-slot (`_eg_qcache_text`,
only ever remembered the LAST query) to a direct-mapped, FNV-1a-keyed,
1024-slot cache (reusing the existing engram_id_hash) — so the
curiosity loop's rotating phrases actually hit the cache instead of
evicting each other every call. Same "pointer owned by the cache, not
freed by caller" contract as before, just per-slot instead of global.
3. Beam cap on the layer-1 spreading-activation BFS (new
engram_activate_beam(), tunable via ENGRAM_ACTIVATE_BEAM, default 128).
The FIFO frontier is processed in hop-level batches (entries sharing
.hops are provably contiguous — see the code comment); when a level
exceeds the beam width, only the top-`beam` by activation actually
EXPAND. Every node in an oversized level still gets reached[]/best_bg[]
recorded (that happens at enqueue time, one level up) and appears in
the reported/promoted set — the cap bounds associative SPREAD width
only, never recall of what was already found. Kept as a genuine
additional bound even though the adjacency index + qgate + fan effect
already mitigate #105's original "hub-node explosion" failure mode for
a different reason: those prune WHICH targets matter; this bounds
worst-case width regardless.
Everything else in dev's engram_activate — cosq/e_eff, the qgate rescale,
the fan effect, eg_edge_eff_weight, the M8 HNSW vindex seed discovery from
the #109 reconciliation earlier this session — is untouched.
VERIFIED (nsbx sandbox only, live :8742/:7770 never touched): cc -std=c11
-O2 clean build; booted in an isolated sandbox against a real cloned
production snapshot (13,424 nodes / 37,656 edges); ran 5 activate() calls
across rotating queries at depth 3, including the same query issued twice
non-consecutively (2nd hit landed at 476ms vs the 1st at 483ms — consistent
with a cache hit once Ollama's own warm-model latency is accounted for; no
crash, correct varied result counts (367-2610 nodes) each call; act-stats
JSON read correctly throughout.
Built on top of the M8/#109 reconciliation (bacaf3d, merged to dev as
#109) — dev's current HEAD at the time of this commit.
Lands feat/reframe-region-setop (PR #109: native set-based reframe_region,
decorator-as-seam @route port, teacher-summon, and the M8.1 activate-latency
work — lazy-memoized cosq via eg_cosq_at + engram_vindex HNSW-accelerated
seed discovery + vindex_harvest_from_store/vindex_bench oracle) onto dev's
actual current HEAD, plus engram-tiered-storage's still-unique test suite.
RECONCILING #109 WITH engram-tiered-storage (M4-M10 HNSW/geometry/reason/
verify work): not a two-way merge. engram_vindex.c's HNSW core (search_layer/
select_neighbors/prune_links/insert) is BYTE-IDENTICAL between the two
branches; #109's copy is a strict superset (adds vindex_harvest_from_store,
used by vindex_bench.c's brute-force-vs-HNSW oracle). engram_reason.c and
engram_verify.c are also byte-identical. #109's own branch point already
carried engram-tiered-storage's M4-M10 lineage forward, so there was nothing
left to merge into #109 for those files. The one thing engram-tiered-storage
had that #109's tree dropped: its full test suite (test_vindex.c,
test_geometry.c, test_reason.c, test_verify.c, test_m7_traversal.c, the
interoception P0-P5 tests, bufpool/compaction tests, and their run_*.sh
harnesses) — ported over here unchanged.
WHY THIS NEEDED HAND RECONCILIATION, NOT A MECHANICAL MERGE: #109's branch
forked from dev on 2026-08-14 15:40 (before restructure-adjacent history
diverged the file's merge-base for `git merge` — it presented as an add/add
conflict). A straight two-dot diff (dev tip -> PR tip) applied cleanly, but
it silently reverted THREE dev fixes landed on 2026-08-14/15, after the
branch point, that the PR's diff had no way to know about:
1. qgate rescale (2026-08-14 self-review): PR's lazy eg_cosq_at rewrite of
the query-aware propagation gate dropped the shift-and-floor rescale
about ENGRAM_EMBED_S0 (measured: unrelated-pair median 0.562->raw gate
0.67, i.e. "a small tax, not a gate"). Restored the rescale, wrapped
around the lazy accessor -- the PR's actual improvement (WHEN cosq[oi]
is computed) is orthogonal to WHAT it gates on and both are kept.
2. Eviction cause decomposition (2026-08-14 self-review): dev decomposes
wm_evicted into evict_floor/evict_cap/evict_bll so WM churn is
diagnosable (identity: evicted == floor+cap+bll+dup_wm+dup_wm_global).
PR's tree predates this and dropped all three counters + their JSON
stats fields. Restored declarations, all 4 direct increment sites, the
eg_wm_carry_over bll increment, and the act-stats JSON fields --
alongside (not instead of) the PR's own P4 afferent / API-reshape
counters already in that same struct/JSON.
3. Hebbian link-formation selection (2026-08-15 self-review, TODAY): dev
selects the STRONGEST qualifying candidate for consolidation each call;
PR's tree predates this and reverted to hash-slot order (arbitrary wrt
association strength) for edge formation -- the one path that writes
PERMANENT structure. Restored the strongest-candidate while-loop,
keeping the PR's own genuine improvement at that site
(engram_adj_on_edge_added incremental-index append instead of a bare
adj_dirty=1 full-rebuild flag).
engram/src/server.el's 3-way conflicts (autoconnect_on/ise_offgraph_on env
flags, /api/nodes connected-count in responses) were pure additive: dev's
side was empty, PR's side added the feature. Took PR's side whole.
VERIFIED (nsbx sandbox only, live :8742/:7770 never touched):
- cc -std=c11 -O2, clean link against the real engram/src/server.el via
elc, zero errors.
- vindex_bench (built standalone, read-only harvest) against the real
production store clone (13,671 embedded nodes, 768-dim nomic-embed-text):
recall@10 = 1.0000 at ef 64/128/200; HNSW search 0.28-0.79ms/query vs
2.03ms/query brute-force oracle (2.6x-7.2x). HNSW build itself: 46.5s
for the full 13,671-node set -- see the flagged risk below.
- Booted the reconciled binary in an isolated nsbx sandbox (:8905, cloned
snapshot of the live store, 13,424 nodes / 37,656 edges) and called
/api/activate for real: first call after boot 41.5s (pays the one-time
HNSW build inline -- matches the standalone bench), second/third calls
356ms/605ms, no crash, correct results, act-stats JSON (including the
restored evict_floor/cap/bll fields) reads correctly.
KNOWN RISK TO FLAG BEFORE ANY LIVE CUTOVER (not fixed here; out of scope for
this dev-only land per instructions not to touch :8742/:7770): eg_vindex_sync
builds the HNSW index synchronously, inline, on the first engram_activate()
call after every process start (or index invalidation). On the real node
count that is a ~46s blocking stall on a single-threaded server -- the first
request after every restart (or its concurrent siblings) waits the full
build. Recommend a background/incremental build (or a bounded per-call build
budget) before this ever reaches the live daemon. See PR description / final
report for the fuller writeup.
ingest and transduce are complements, not synonyms: ingest is the conscious,
deliberate act of pointing at a source (ingest_file/dir/url/llm/stream stay
named exactly that); transduce is the automatic, invisible mechanism inside
it that converts extracted surface content into geometry (renamed
build_prose/build_structured -> transduce_prose/transduce_structured, the
functions that actually turn raw text into a node+edge manifold).
Real bug found and fixed along the way: the final /api/load-merge response
was never checked for an error. A total failure (bad auth, network down,
anything) silently reported nodes_added:0/edges_added:0 — indistinguishable
from a benign 'everything was already known' outcome. Verified live: with a
wrong key, the tool now honestly returns {"error":"load-merge failed:
unauthorized",...} instead of a misleading zero.
Also dropped a CRUD-verb smell: the per-decision println said CREATE (a
database-log verb for something that hasn't actually been written to the
server yet — it's a local, tentative decision pending the batch merge).
Renamed to FORM. The dead-code eg_create_node (defined, never called)
renamed to eg_crystallize_node and annotated honestly as unused, since if
it's ever wired up it represents the real server-confirmed write, unlike
the local FORM guess.
Not yet re-verified end-to-end against a real successful write: the
ingest-test sandbox (nsbx up ingest-test) is itself currently broken —
it prints a green "ready" banner after its own readiness check fails,
and nothing is actually listening. Filed separately; not in scope here.
AGENTS.md: root-level guide to the repo — which of the 8 el_runtime.c
copies is the one canonical, authored source (lang/releases/v1.0.0-20260501,
despite the misleading 'releases/' name) vs. lagging forks/build artifacts,
build commands, and session protocol.
engram/spec/architecture-hardening.design.md: terse engineering anchor for
the 2026-08-14 hardening vision (one calculus over the geometry, core +
ephemeral ring, persistence earned by salience, incarnation model) —
indexes the fuller whitepaper + Neuron artifact 2b8078cf rather than
restating them.
engram/spec/engram-db-tooling-design.md: high-level design for engram DB
tooling (geometry-native browse/query/ops surface over the existing
vantage-read/write/relate/supersede API).
Deliberately leaves out of this commit: the uncommitted el_runtime.c/h +
codegen.el float-arithmetic-codegen diff in this worktree, which appears
to overlap with (or supersede) the fix already preserved via PR #104 —
needs manual reconciliation rather than a second competing PR. Also
leaves out lang/.promote-backup-floatfix/ (a local backup snapshot,
confirms that float-fix work is mid-promotion here), assorted .DS_Store
files, engram/dist/engram.* backup binaries, and lang/dist backup
binaries — none of it source.
939-line Swift I/O organ (mic/camera capture, speaker playback via
AVFoundation/CoreAudio), own-core LPC voice synthesis/imitation,
consent-gating, and full-duplex barge-in conversation — closing the
hear -> understand -> speak loop entirely on-device.
.gitignore in this dir already excludes bin/ (build output), out/
(captured media), and .consent.json/.resume.json (local runtime state),
so only src + README + .gitignore are committed here.
speech.el: formant/glottal integer DSP synthesis + voice-analyze-by-
imitation. voice-profile.el / voice-ingest.el: voice-profile plumbing.
accent.el: British-RP as an ingested transform-geometry (explicitly marked
provisional/citation-pending by its own comments). organ-read.el:
engram read-through for the speech organ. Includes demo/test drivers and
non-personal reference data (British-RP phonetics/lexicon derived data,
a public-domain LibriVox RP reference recording).
Deliberately excludes elp/data/live/ (raw recorded voice + face-photo
samples of the repo owner) and the will-*.{json,psv} derived voiceprint
files — personal biometric data that shouldn't be committed to a shared
repo without an explicit decision from the owner. Also excludes this
worktree's elp/src/surface-profile.el, which diverges from the copy in
other worktrees (agent-aaf04b0a9714c4070, main) — needs manual
reconciliation before landing, left out here to avoid silently picking a
version.
audio-surface.el / image-surface.el: own-core additive-synthesis WAV and
raster-PNG renderers (integer-only DSP, since EL has no floats), rendered
from learned engram signatures via a pluggable surface-profile
abstraction (surface-profile.el). audio-demo.el / image-demo.el are
drivers. NOTE: demo files hardcode absolute paths to this worktree's own
directory — will need a path fixup before landing.
elp/projector/ is a Python package the author's own README marks as
"STAGING/PROOF-OF-SHAPE — not the deliverable", superseded by the native
.el surface-profile work above; kept as a validated architecture proof.
Generated output (elp/faculty/{out,sig}, elp/projector/out,
__pycache__) intentionally excluded.
Adds an O(1) "seen" bitmap so lazily-embedded older nodes get picked up
incrementally instead of only on a full rebuild (embed-gap #20).
Replaces engram_activate's O(N*D) cosine prescan with a lazy-memoized
cosine cache (eg_cosq_at), proven bit-identical to the old path.
Extracts a clean vindex_harvest_from_store primitive (read-only vector
harvest, careful malloc/ownership/error-path handling) reused by both
index-build and the new vindex_bench.c — a read-only proof harness
comparing brute-force vs HNSW recall/latency on both the real store and
synthetic data.
.nsbx-env intentionally excluded — local sandbox config (ports, paths,
dev-only placeholder key), not checked in.
LTP/LTD-style belief grounding propagated along graph edges, with
union-find independence-guarded corroboration. Package: core C algorithm
(gep_core.h), a self-contained deterministic proof harness with recorded
output, staged runtime integration, and gated .el patches for the beat
hook and HTTP route.
Per the author's own LEDGER.md: built + proven on a clone, GATED pending
the engine/HNSW cutover — not wired into the live beat or routes.
Preserved here as a spec/reference artifact, not a request to merge into
the live path.
Cancellation-token control channel checked at every step boundary lets a
coordinator PAUSE/RESUME/REDIRECT/KILL a running worker mid-task instead of
waiting for the whole (possibly wrong) plan to finish. Bounded purviews
mean no half-committed state to unwind on interrupt. Includes a proof
harness (proof.el, run.sh) comparing a broken non-interruptible worker
against the new one under identical kill/redirect/pause timing.
Distinct from the already-preserved swarm-ccr orchestrator (fan-out/
converge dispatch): this is single-worker interruptibility, a
complementary mechanism, not a duplicate.
Float + previously fell through to string concat (segfault); -, *, /, %
operated on raw IEEE-754 bit patterns as integers (garbage results). Floats
are now tracked via a __float_names typed-binding set (parallel to the
existing int-tracking scheme) and arithmetic is emitted as real C double
ops.
Also fixes math_log, which was wrongly aliased to natural log (duplicating
math_ln) — now uses log10 — and adds the missing <math.h> include. Rebuilt
elc binary included.
Adds engram_assert_json — a grounded "assertion envelope" primitive for a
realizer/op_assert seam (per backlog bl-53/#57) — plus purview-scoped
mutation wrappers engram_node_full_in/engram_connect_in, which refuse
non-default purviews rather than silently mutating the live store. Threads
through el_seed.c/h wrappers and the codegen.el arity table per the
project's existing C-builtin recipe.
Also rewrites lang/AGENTS.md build docs with verified (2026-08-15) findings
that el_seed.c does not compile standalone.
Source-polymorphic ingest(source) primitive: extracts content faithfully
from a directory/file/url/llm-query/structured-primitive-set/stream,
decomposes it into a discrete multi-node graph manifold (nodes + internal
edges, never a single blob), and merges it into the engram geometry with
dedup (search + exact/cosine match), provenance, grounding-level, and
stewardship-class tagging from the moment of entry.
Pure HTTP client of the engram server (links only el_runtime.c, never
el_seed.c/the engine directly). Tested against a live nsbx sandbox engram
clone (127.0.0.1:8903) with real writes confirmed via /api/stats
(node_count 3201 / edge_count 6601).
Excludes ingest/build/ — local compiler scratch output (binaries, .c
codegen, .err logs), not source.
The link-formation scan walked candidate slots ascending and stopped at
ENGRAM_HEBB_LINK_PER_CALL (2). Slot index is a hash of the node id pair, so
whenever more than two candidates cleared LINK_MIN in the same call, the two
consolidated were the two with the lowest hash and a stronger association
waited - indefinitely, since the scan restarts from slot 0 every call while
the leader decays at ENGRAM_HEBB_DECAY.
Measured 08-13..08-15: hebb_cand_max peaked at 0.4963, 3.3x LINK_MIN, during
a ~14h stretch of continuous qualification at the 2/call cap.
Same defect the 2026-08-02 review named and fixed for breakthrough weights
(index order is not a cognitive criterion), never carried across to the one
path that writes permanent structure - and there is no pruning path, so
growth is one-way. Selection pressure matters most where the result is
irreversible.
No-op when <=2 candidates qualify; picks the best when more do.
Add the universal engram mutation as ONE operation: isolate a region
(cosine + adjacency) -> supersede it as a set (immutable region-tombstone,
originals retained, engram_forget never used) -> insert the new manifold as a
set -> rebind edges by cosine -> one atomic persist. Single-node write and
supersede are the degenerate n=1 case of the same reframe_core path, not a
separate CRUD path. Keystones kn-efeb4a5b / kn-5b606390 are write-protected.
Purely additive: routes POST /api/reframe, /api/write, /api/supersede.
Verified on an isolated clone of the JSON-snapshot engine (set-replace, n=1,
no-regression, keystones, durable reboot); compile-verified clean against the
cognition multi-TU build. NOT deployed — prod :8742 frozen; blue-verify on the
cognition/egm engine required before any cut.
bigmerge
will.anderson