# Capacity gate — pause/catch-up/restart with exact reconciliation (R26-11) Run `cap-20260811T181546Z-98337`, ap-southeast-1 (Singapore), 2026-08-11/12. Build: wrapper-verified sha `5a1c3b24…` (streams) / `8f184342…` (awsbench) against the campaign manifest. Workload: 32 streams, concurrency 64, 10 × 1 KiB records/request, FRAME_COMPRESS=1, `LIMIT_RECS_PER_SEC=100000` (campaign posture — the ordinary limiter deliberately out of the way), compute-1g memory profile, INITIAL_SHARDS=4. Driver: `bench/soak/capacity-gate.sh`. Artifacts: `$SOAK_HOME/results/cap-20260811T181546Z-98337/`. ## Verdict: PASS (criterion B), with exact zero-loss reconciliation | Check | Result | |---|---| | Exact op-ledger reconciliation | **OK — acked 11,248,940 == durable 11,248,940 == walked 11,248,940**, 32 streams, 1,124,894 acked ops each exactly-once with all batch positions, 0 ambiguous, 0 problems | | B: backlog bounded | **peak 34.4 MB instance / 11.6 MB max-shard vs 512 MiB / 256 MiB caps** — never approached; latch never engaged; server shed = 0; rate-limit refusals = 0 (typed counters, not inference) | | A: catch-up ≥ 1.25× ingest | formally not established by the windowed counter measurement (1.16 across a counter-resetting restart); the ledger series shows 34.4 MB retired in ≤ 100 s against ~188 KB/s steady ingest — an implied ratio > 2 — see "measurement honesty" | | Restart at max backlog | process replaced at ledger peak; **serving ≤ 1 s after the 51 s redeploy completed** (bound 300 s) — durable maintenance load + WAL replay inside the deploy window | | Recovery window | ledger 0, no latches, **0.8 s** after ramp end (fixed-window check) | | Errors | **0** across 1.6M+ requests | ## What the run measured - **Steady state (90 min):** ~278 req/s ≈ 2,780 rec/s accepted; **exact frame-byte ingest 187.6 KB/s** (the all-'x' payloads compress ~15:1 — the same unit correction that retired the 9.4% artifact, visible live). Ledger oscillated 2–13 MB: absorption kept pace continuously, progress clock never aged. - **Pause (183 s):** ledger grew linearly 8.9 → 34.4 MB, exactly ingest × time. No shed (nowhere near caps at this frame intensity). - **Resume + restart:** the new process loaded the durable ledger and the absorber retired the full backlog within ~100 s. - **Integrity through a fencing storm:** the restart leg created a version overlap — the generator kept the retired version's domain, which the platform answered with 503s (484,298 of them) once the old process was fenced/stopped. Every one was filed as a definitive rejection in the op ledger, and the exact walk still balanced to zero: no acked record missing, none duplicated, through fencing, restart, and edge 503s. This is the strongest integrity evidence the project has produced. ## Measurement honesty (read before quoting) 1. **PASS came via criterion B**, and criterion B was not stressed: at 187 KB/s compressed-frame intensity, a 3-minute pause builds 34 MB — 6.6% of the shard cap. The bound's shed behavior is proven deterministically (DST + the R25-H part-1 gate at 8/8), not by this run. The stressing field case is an INCOMPRESSIBLE workload: the same record rate would ingest ~2.9 MB/s of frame bytes, a 3-minute pause would build ~520 MB, and the caps + typed shed would be the thing holding the line. That run is the natural next experiment now that the harness can attribute every refusal. 2. **Criterion A's number (1.16) is a measurement artifact**, not a capability statement: the restart resets the process-local cumulative counters mid-catch-up, and the evaluator drops the negative delta, so the window mostly covers the paused (flat) segment. The durable ledger series — 34.4 MB to zero in ≤ 100 s against 188 KB/s live ingest — implies a catch-up ratio well above 2×. Durable counters (or restart-surviving totals) would make A formally measurable across restarts. 3. **The post-restart load collapsed** (~25 req/s) because the generator targets the version-scoped preview domain captured at deploy time; the restart mints a new version. So the catch-up ran against reduced live load. Harness fix for next time: point generators at the service-stable domain, or redeploy/repoint the generator after the restart leg. ## Disposition - CHAOS-5 is closed as originally stated: the 9.4% figure was the payload-vs-frame unit artifact; with exact units, absorption keeps pace at this profile and the durable bound + typed shed guard the regimes where it cannot. - The supported-capacity number for INCOMPRESSIBLE frame intensity remains to be measured (item 1 above) — the harness is now capable of measuring it honestly. - Teardown verified: services, bucket, project deleted; receipt retired.