CS07A — Helpfulness benchmark harness

Date: 2026-10-11. Status: LOCAL_DONE. Scope: harness + local safe-rate runs only — no route/schema/migration change, no VPS/APK, no staging run. Head: maintained dev. Plan: phased plan. Rules: B-BR-CS07.

Harness (adapters/helpfulness_bench_test.go)

Exact application paths (Cast/TallyFor) on a disposable database with a separate 50-op warmup and fixed inner loops (200 ops, 8 workers): tally-read, cast-fresh-voter (contended tally row), flip-same-voter. Reports p50/p95/p99/max + wall ops/s. Tally SQL mirror with drift test against helpfulness.sql (RST-07 style). Shared seeding helpers generalized from *testing.T to testing.TB (compatible, no behavior change).

Local numbers (disposable PostGIS, 8 workers, 200 iters, warmed)

Two consecutive runs, stable:

• tally-read: p50 0.07ms, p95 0.28ms, wall ~34k op/s.
• cast-fresh-voter: p50 3.6ms, p95 8.9ms, wall ~1.9k op/s.
• flip-same-voter: p50 2.9ms, p95 15.2ms, wall ~1.5k op/s.

p99 spikes (4–19ms) on tiny windows are warmup/scheduling noise; staging 120s windows smooth them. Local wall rates prove the harness, never capacity: no max-user claim, ever.

Frozen staging protocol (CS07B, authorized window)

5 independent 120s runs per scenario with separate warmup at safe rates; per-run kernel/runtime evidence attached; guarded 5 req/s ceiling frozen only when the selected tail p95 holds with ≥5% headroom; runs stop below sampled pool/row/storage thresholds with wall-clock throughput recorded. Ingestion stays narrow with rows/day within twice the T-30 median; baselines on volunteers/public data.

Checks

Harness + drift test green twice; existing helpfulness integration re-green after the TB change; go vet (integration tags), gofmt -l, git diff --check clean; scan-secrets.sh passed.

Next microtask: CS07B — staging guard runs (authorized window).