Files
el/engram/test/bench_discrimination.c
T
bigmerge bacaf3d39c
El SDK CI - dev / build-and-test (pull_request) Failing after 4m49s
engram: reconcile M8 HNSW vindex (#109) onto current dev, restore 3 fixes the branch predated
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.
2026-08-15 16:46:44 -05:00

150 lines
7.4 KiB
C

/* bench_discrimination.c — M9 REFINEMENT bench: measures whether mean-centering
* the anisotropic nomic-embed-text space sharpens the §5 geometry operators on
* REAL data. Read-only over a COPY of the live store (never the live file).
*
* usage: bench_discrimination [store.egm]
* (or set ENGRAM_BENCH_STORE). If no store is given/openable it prints
* SKIP and exits 0 — so it is safe in CI without live data.
*
* It picks two semantically distinct cohorts by keyword (domain A vs domain B),
* computes the global mean over the embed-eligible set (via engram_geo_mean_build
* — the same offset the descriptor uses), then reports BEFORE (raw unit space)
* vs AFTER (mean-centered space):
* - cross-centroid cosine (lower = better separated)
* - cross-centroid Euclid dist (translation-invariant: a control)
* - intra-cohesion per domain (member cos to own centroid)
* - overlap operator (cross_cos / sqrt(intraA*intraB): ~1 = domains
* indistinguishable, ~0 = cleanly separated)
* - angular separation ratio z (centroid angle / summed angular spread)
* - mean pairwise cosine sample (the anisotropy headline; ~0.55 raw -> ~0 ctr)
*
* Pure C11; links engram_store.c + engram_geometry.c; -lm.
*/
#include "engram_store.h"
#include "engram_geometry.h"
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <strings.h>
#include <math.h>
#define CAP_DOMAIN 400
#define CAP_SAMPLE 800
typedef struct { float** v; int n, cap, dim; } VecSet;
static void vs_init(VecSet* s){ s->v=NULL; s->n=0; s->cap=0; s->dim=0; }
static void vs_push(VecSet* s, const float* e, int dim, int cap){
if(s->n>=cap) return;
if(s->dim==0) s->dim=dim;
if(s->n==s->cap){ int nc=s->cap?s->cap*2:64; s->v=realloc(s->v,(size_t)nc*sizeof*s->v); s->cap=nc; }
float* c=malloc((size_t)dim*sizeof(float));
double nn=0; for(int d=0;d<dim;d++) nn+=(double)e[d]*e[d]; nn=sqrt(nn);
if(nn<1e-12){ free(c); return; }
for(int d=0;d<dim;d++) c[d]=(float)(e[d]/nn); /* L2-normalized copy */
s->v[s->n++]=c;
}
static void vs_free(VecSet* s){ for(int i=0;i<s->n;i++) free(s->v[i]); free(s->v); }
typedef struct { VecSet A, B, S; long idx; } Coh;
static int has(const char* h, const char* n){ return h && strcasestr(h,n)!=NULL; }
static void cb(const StoreNode* n, void* ctx){
Coh* c=ctx;
if(!(n->emb && n->emb_dim>0)) return;
/* every 5th embedded node -> isotropy sample */
if((c->idx++ % 5)==0) vs_push(&c->S, n->emb, n->emb_dim, CAP_SAMPLE);
const char* t=n->content; const char* g=n->tags;
int A = has(t,"quantiz")||has(g,"quantiz")||has(t,"lorablation")||has(t,"70B")||has(t,"LoRA merge");
int B = has(t,"kubernetes")||has(t,"terraform")||has(t,"argo")||has(g,"infrastructure")||has(t,"vault")||has(t,"cloudflare");
if(A && !B) vs_push(&c->A, n->emb, n->emb_dim, CAP_DOMAIN);
else if(B && !A) vs_push(&c->B, n->emb, n->emb_dim, CAP_DOMAIN);
}
/* mean of a VecSet into out (dim doubles). */
static void mean_of(const VecSet* s, const float* gm, double* out){
int dim=s->dim; for(int d=0;d<dim;d++) out[d]=0;
for(int i=0;i<s->n;i++) for(int d=0;d<dim;d++) out[d]+=(double)s->v[i][d]-(gm?gm[d]:0.0);
if(s->n) for(int d=0;d<dim;d++) out[d]/=s->n;
}
static double dnorm(const double* a, int dim){ double s=0; for(int d=0;d<dim;d++) s+=a[d]*a[d]; return sqrt(s); }
static double dcos(const double* a, const double* b, int dim){
double na=dnorm(a,dim), nb=dnorm(b,dim); if(na<1e-12||nb<1e-12) return 0;
double s=0; for(int d=0;d<dim;d++) s+=a[d]*b[d]; double c=s/(na*nb);
if(c>1)c=1; if(c<-1)c=-1; return c;
}
static double deuclid(const double* a, const double* b, int dim){
double s=0; for(int d=0;d<dim;d++){ double x=a[d]-b[d]; s+=x*x; } return sqrt(s);
}
/* mean cosine of members (minus gm) to centroid c (already gm-subtracted). */
static double cohesion(const VecSet* s, const float* gm, const double* c){
int dim=s->dim; double nc=dnorm(c,dim); if(nc<1e-12||s->n==0) return 0;
double acc=0; for(int i=0;i<s->n;i++){
double dot=0, nv=0;
for(int d=0;d<dim;d++){ double v=(double)s->v[i][d]-(gm?gm[d]:0.0); dot+=v*c[d]; nv+=v*v; }
nv=sqrt(nv); if(nv<1e-12) continue; double cc=dot/(nv*nc);
if(cc>1)cc=1; if(cc<-1)cc=-1; acc+=cc;
}
return acc/s->n;
}
/* mean pairwise cosine over a sample (isotropy metric). */
static double mean_pairwise_cos(const VecSet* s, const float* gm){
int dim=s->dim; if(s->n<2) return 0; double acc=0; long np=0;
for(int i=0;i<s->n;i++) for(int j=i+1;j<s->n;j++){
double dot=0, na=0, nb=0;
for(int d=0;d<dim;d++){ double a=(double)s->v[i][d]-(gm?gm[d]:0.0), b=(double)s->v[j][d]-(gm?gm[d]:0.0);
dot+=a*b; na+=a*a; nb+=b*b; }
na=sqrt(na); nb=sqrt(nb); if(na<1e-12||nb<1e-12) continue;
double c=dot/(na*nb); if(c>1)c=1; if(c<-1)c=-1; acc+=c; np++;
}
return np? acc/np : 0;
}
static void report(const char* label, Coh* c, const float* gm){
int dim=c->A.dim; double* ca=malloc((size_t)dim*sizeof(double)); double* cb=malloc((size_t)dim*sizeof(double));
mean_of(&c->A, gm, ca); mean_of(&c->B, gm, cb);
double xcos=dcos(ca,cb,dim), xeuc=deuclid(ca,cb,dim);
double cohA=cohesion(&c->A,gm,ca), cohB=cohesion(&c->B,gm,cb);
double overlap = (cohA>0&&cohB>0)? xcos/sqrt(cohA*cohB) : xcos;
double theta = acos(xcos<-1?-1:(xcos>1?1:xcos));
double sig = acos(cohA<-1?-1:(cohA>1?1:cohA)) + acos(cohB<-1?-1:(cohB>1?1:cohB));
double z = (sig>1e-9)? theta/sig : 0;
double mpc = mean_pairwise_cos(&c->S, gm);
printf(" [%s]\n", label);
printf(" cross-centroid cosine = %+.4f (lower = better separated)\n", xcos);
printf(" cross-centroid Euclid = %.4f (translation-invariant control)\n", xeuc);
printf(" intra-cohesion A / B = %.4f / %.4f\n", cohA, cohB);
printf(" OVERLAP operator = %.4f (~1 = indistinguishable, ~0 = clean)\n", overlap);
printf(" angular separation z = %.3f (centroid-angle / summed spread; >1 = separated)\n", z);
printf(" mean pairwise cosine = %+.4f (isotropy: ~0.55 anisotropic -> ~0 isotropic)\n", mpc);
free(ca); free(cb);
}
int main(int argc, char** argv){
const char* path = (argc>1)? argv[1] : getenv("ENGRAM_BENCH_STORE");
if(!path){ printf("SKIP: no store path (arg or ENGRAM_BENCH_STORE)\n"); return 0; }
EngramPagedStore* st=store_open(path);
if(!st){ printf("SKIP: could not open %s\n", path); return 0; }
Coh c; vs_init(&c.A); vs_init(&c.B); vs_init(&c.S); c.idx=0;
store_scan_nodes(st, cb, &c);
printf("=== two-domain discrimination bench (real store copy) ===\n");
printf("domain A (quantization) n=%d ; domain B (infrastructure) n=%d ; sample n=%d ; dim=%d\n",
c.A.n, c.B.n, c.S.n, c.A.dim);
if(c.A.n<3 || c.B.n<3){ printf("SKIP: a cohort is too small to be meaningful\n");
vs_free(&c.A); vs_free(&c.B); vs_free(&c.S); store_close(st); return 0; }
GeoMeanCache* mc=engram_geo_mean_build(st);
const float* gm=engram_geo_mean_vec(mc);
printf("global-mean cache: dim=%d over %llu embedded nodes\n\n",
engram_geo_mean_dim(mc), (unsigned long long)engram_geo_mean_count(mc));
printf("BEFORE (raw anisotropic unit space):\n");
report("RAW", &c, NULL);
printf("\nAFTER (mean-centered isotropic space):\n");
report("CENTERED", &c, gm);
engram_geo_mean_free(mc);
vs_free(&c.A); vs_free(&c.B); vs_free(&c.S);
store_close(st);
return 0;
}