feat(phase-55): Performance Truth — split setup vs query, honest percentiles

[RichardHightower]

Summary

Implements v3.1 Phase 55 (Performance Truth). No new retrieval capabilities — the recorded numbers now match what the steps actually do.

Phase 54 (#32) and the v3.1 spec (#31) are already on main.

Root cause of the 64.6s TOC p50

perf_bench timed build_toc_segment (MockSummarizer rollup of 240 events) under the step name toc. Two get_toc_node lookups were in the same timer. p90/p99 were interpolated from 3 samples.

What this makes true

Plan	Claim	Now
55-01	TOC query is navigation, not rollup	*.toc = three get_toc_node lookups. Rollup is *.toc_build. single.toc warm p50 = 0.13 ms (n=30, 240 events). single.toc_build = 76.7 s — that is the retired number, honestly labeled.
55-01	Vector init vs query	vector_model_load (156 ms cached), vector_index (12.6 s), vector (query embed + HNSW).
55-02	Percentiles survive scrutiny	Default 30 query iterations. p90 omitted unless n≥10; p99 unless n≥30. Setup steps report min/median/max only.
55-02	Cold vs warm are structural	Warm: one setup, one discarded warmup, N query samples. Cold: new store per iteration.

Execution evidence

Committed crates/e2e-tests/benchmarks/latest.json (schema 2, 2026-08-30, linux/x86_64, debug cargo run, 240 synthetic events, 30 warm query samples). docs/benchmarks.md records corpus, hardware, profile, and caveats.

Unit tests in perf_bench.rs cover percentile gating, including that p99 of 3 samples is an interpolation artifact.

Not in this PR

• Release-profile vector numbers (debug Candle query-embed is ~4 s; do not quote as production HNSW).
• Speeding up toc_build (ingest-time rollup). It is now labeled, not mis-sold as query.
• Phases 56–58.