Throughput Live-Proof Session Log (2026-09-05)

v1.28.58 “Throughput” — the milestone’s live proof record, per the execution prompt: bench measured runs (3×, desktop), the same-seed determinism pair, the three /metrics captures around a parallel burst, and the CRA drill baseline. All against a COPY instance — the live deployment was untouched.

Environment

• Copy instance: BIND_PORT=18765, fresh scratch DB, opaque-token auth (AUTH_TOKEN_FILE, 0600). Release build (cargo build --release --features bench --bin brain-server --bin bench).
• Corpus: 1 000 synthetic docs (BENCH_SCALES=1000), ingest ≈ 1 050–2 500 docs/s on this desktop box.
• Rate-limit arithmetic observed: the per-IP limiter is 10 000 req/min; a default-scales run (1k+5k+10k cumulative ingest) trips it — every measurement run here stayed ≈ 2 700 requests, far under the budget. The CI bench-concurrency job (~1 800 requests) has ample margin.

Concurrent bench — three measured runs (desktop, BENCH_CLIENTS=8)

BENCH_CLIENTS=8 BENCH_SEARCHES=200 BENCH_SCALES=1000

Per-client skew across all runs: 8×200 ops, evenly — p50 spread between clients < 1 ms; the deterministic mix means divergence would be server-side queuing, and none was observed.

Ceiling derived: desktop search_p95_ms_ceiling = 60 ms (worst run 23.07 + ~2.5× margin). Jetson stays 150 ms, unmeasured pending a device run (no ARM runner — the known repo CI gap).

Same-seed determinism pair (BENCH_SEED=42, twice)

Structural diff (non-latency columns) between the two merged reports: identical — same total ops ok, same failures, same per-client counts (8 × 200). Latency values are timing physics and vary within noise (p95 spread ≈ 2.7%); the seeded mix makes every breach reproducible.

/metrics captures — before / during / after a parallel burst

Burst: BENCH_CLIENTS=8 BENCH_SEARCHES=800 BENCH_SCALES=10 (6 400 searches, 0 failures). A /health/db scrape preceded capture 1 to populate the WAL snapshot (the PASSIVE-checkpoint PRAGMA lives only there).

Capture 1 — BEFORE:

brain_pool_in_use{domain="global"} 0
brain_pool_idle{domain="global"} 20
brain_pool_timeouts_total 0
brain_busy_errors_total 0
brain_wal_pages_pending{domain="global"} 0

Capture 2 — DURING (8-client search phase in flight):

brain_pool_in_use{domain="global"} 5
brain_pool_idle{domain="global"} 15
brain_pool_timeouts_total 0
brain_busy_errors_total 0

Capture 3 — AFTER:

brain_pool_in_use{domain="global"} 0
brain_pool_idle{domain="global"} 20
brain_pool_timeouts_total 0
brain_busy_errors_total 0

The pool-saturation gauge moves 0 → 5 → 0 with the burst; the counters stay at 0 because nothing waited 30 s for a slot and no write BEGIN burned through busy_timeout — the honest no-contention reading, not a wired-off gauge. brain_wal_pages_pending appears only after a /health/db scrape, per the cold-path design.

CRA drill baseline (tabletop, 2026-09-05T04:40:18Z)

scripts/cra-report-drill.sh — fabricated actively-exploited-vulnerability notice against the current release; filled 24 h template + timing report in dist/cra-drill/ (and /tmp/cra-drill-final/ for this record):

The timings are machine-fast because the tabletop is deterministic shell work — the rehearsal value is the artifact walk (SBOM located, version matrix consulted, template filled, channels named), not the stopwatch. Baseline archived per the DSAR-drill precedent.