Historical report imported 2026-10-10 from docs/experiments/benchmarks/catalog-many-segment-findings.md. Original SHA-256: 630c615b1312c54444fad0d60f503da7ba01a0c2629831af0892c9a5c901752f. Measurements and outcomes retain their stated revisions. Links and local environment labels are normalized; $FABRIC_DATA is the mounted experiment drive. See evidence policy.

CR3 many-Segment query findings

Completed without nomination. Run02 aggregate (local evidence omitted from stage) contains all24 fresh pairs/48 children:96 independently graded canonical complete chains and3072 exact associated measured first pages. Every population used H65536,1024-byte bodies, seed42, Walk, limit1000,32 completions per shape; 1/64Segments,1/4reader threads, clone/shared, plain/count stayed separate. Protocol (local checkout reference docs/experiments/benchmarks/catalog-many-segment-protocol.md; not present at origin/main) and query correction (local checkout reference docs/experiments/benchmarks/catalog-many-segment-query-correction-protocol.md; not present at origin/main) remain frozen. Run01's contains:null rejection stays failed and preserved.

Ratios below are shared/clone, independently derived from original per-child timings.json (CPU/wall,32-sample p95, allocator total/calls/peak-minus-base). All raw-derived metrics equal the24 pair summaries. Ranges span the three fresh pairs; counted timing is not used for CPU/latency acceptance.

Segments/readers/shape	Counted requested bytes	Calls	Incremental peak
1/1/absent	0.999995	0.999907	0.999709
1/4/absent	0.999995	0.999907	1.000191–1.001649
1/1/broad	0.999966	0.999776	0.999980
1/4/broad	0.999966	0.999776	0.970157–1.022647
64/1/absent	0.946979	0.922813	0.363184
64/4/absent	0.946979	0.922814	0.326648–0.328575
64/1/broad	0.999823	0.996042	0.998665
64/4/broad	0.999823	0.996042	0.995326–0.995582

At64Segments, absent cumulative bytes are57,073,312→54,047,232 per32 queries, only5.302093% less, missing the >=10% primary criterion in all six counted pairs. Calls fall344,929→318,305 with1reader (four-reader totals add3 calls). Absent incremental peak falls148,941→54,093bytes with1reader; with4readers, clone562,637–564,273 versus shared183,784–185,068bytes. These are conditional allocation/temporary-peak benefits; Arc-retained versions and whole-process memory are different quantities. Real lifetime controls remain separate evidence.

Segments/readers/shape	Plain CPU	Plain wall	Plain p95
1/1/absent	0.986689–0.998019	0.986654–0.998021	0.965085–1.001507
1/4/absent	1.010417–1.040021	1.011778–1.025837	1.006346–1.070593
1/1/broad	0.987267–0.998793	0.987252–0.999289	0.939223–0.975426
1/4/broad	0.943179–1.023535	0.946132–1.022238	0.919778–1.095757
64/1/absent	0.904371–0.970040	0.905559–0.969058	0.903325–0.985030
64/4/absent	0.883233–1.005258	0.855221–1.000091	0.787868–0.982287
64/1/broad	0.998893–1.002722	0.998799–1.002984	0.995753–1.018530
64/4/broad	0.943269–1.006696	0.941516–1.007171	0.928456–0.996581

Every plain64Segment preservation guard passed; every counted peak guard passed. Two1Segment/4reader plain p95 guards failed: broad pair1 ratio1.095757 and absent pair2 ratio1.070593 exceed1.05. Counted timing changes are instrumentation observations, including a counted broad p95 increase; they establish no speedup. No universal win, full-pipeline claim or production promotion follows.

The24 matrix coordinator receipts all exit0 and record scratch removed, zero swap and zero high/max/OOM events. Matrix elapsed814.309838s; whole-job cgroup peak range1,126,408,192–1,197,789,184bytes (1074.23–1142.30MiB), sampled owned scratch max331,315,935bytes (315.97MiB). Freeze02 exit0 took4.040283s; summary receipt (local evidence omitted from stage) exit0 took1.003657s at45.34MiB cgroup peak. These three components total819.353778s, excluding earlier failures/controls. Read-only retained-run inventory is73,310,244 unique file bytes/74,633,216 allocated bytes; evidence is under repository/home, build/temp fixtures under the mounted data drive.

The summary is an artifact/receipt audit: it recomputed original timing metrics, checked24 cells and frozen binary SHA fields, historical oracle verdicts, cleanup and associations, and gzip-read/hash-checked every canonical first-page object. It did not rerun semantic grading or rehash frozen executable archives, source files or every continuation page. Full-chain semantic grading ran during each child collection; selected executable and driver/verifier hashes were checked during pair admission. All24 pair identity receipts record one raw-ledger SHA 6ff69cb65c2bc45fcc6e63e0377e1524664fac0fc3894b2844c4374cedaf18c9. Nomination=false means completed collection with failed acceptance, not missing populations. CR2/CQ2 and historical failed gates are unaffected by this finding.