Serving load requirement evidence

This matrix covers all 58 user stories in the approved specifications. Story numbers refer to the numbered lists in environment, traffic, and gates. Deterministic checks prove controls and failure behavior; disposable runs prove actual preparation, HTTP traffic, measured performance, and teardown. See the handoff for commands, outcomes, and qualification limits.

Environment: 16 stories

Stories	Implementation	Observable evidence
1–2, 11	Controller, owned Compose, Make commands	Full invocation, allocated loopback endpoints, unique run resources, ownership refusal test
3–5	Preparation, dedicated profile	Seed/cutoff manifests, verified row receipts, four independently scoped heads, announced fixture scale
6–8	Normal worker publication and all 44 HTTP readiness checks in controller	Actual four-scope PostgreSQL setup and public serving; wrong-head contract tests
9, 15	Separate Docker stages, optional load extra, pytest plugin exclusion	Delivery contracts, ordinary socket isolation test, Python runtime smoke checks
10, 16	Resource preflight, inspected caps, finite diagnostics and aggregate observations	Full run resource/version evidence; setup failure, five preflight command failures with safe operation/exit metadata, and required-report lifecycle tests
12–14	Setup/request/drain/total/cleanup budgets, signal handling, verified ownership and recovery	Completed run teardown; foreign-volume refusal, setup failure cleanup, failed recovery tests

Serving traffic: 20 stories

Stories	Implementation	Observable evidence
1–5, 9	Workload policy, FastHttpUser driver	Seeded selection, 44 finite labels, scope/family distribution unit tests, full reference traffic
6–8	Source eligibility/category metadata plus exact published-head lane digests	Real API contract tests: wrong scope, rank, score, provenance, corrupt eligibility/category, declared empty/stale
10–13	Fixed reference/stepped/sustained populations, closed-loop phase offsets, achieved valid RPS	Profile tests, actual user pacing regression probe, virtual-time phase checks, measured reference report
14–15	Finite outcomes, no retry/redirect, absolute two-second timeout, bounded drain	Actual HTTP admission/timeout smoke, controlled start-cohort and overload-policy tests
16–17	Fixed bounded dashboard and headless runner	Real Flask mutation hooks reject outside host/population/reset/restart; stop invalidates run; runtime smoke
18–19	Sanitization before event listeners, fixed labels, finite counters, CPU saturation detection	Actual request-event/error sanitization and timeout smoke; report resource observations and original-counter tests
20	Locked Locust 2.46.6 in isolated optional runtime	Real client smoke on macOS ARM64 Python 3.12, 3.13, 3.14; Linux ARM64 generator-image smoke

Performance gates: 22 stories

Stories	Implementation	Observable evidence
1–2, 21	Make reference/capacity commands, manual serialized CI workflow	Full headless invocation and existing repository gates; delivery tests preserve manual rollout, anonymous job-private Docker configuration, and recovery/upload; client configuration tests preserve context/plugins and personal settings
3–5	Half-open request-start windows, post-population settling, full fixed denominator	Warm-up/drain attribution, idle-tail and virtual-time phase checks
6–8, 13	Measurement policy: exact counters, required buckets/samples, rate and completeness	154 ms rounding regression, exact 90/150 boundaries, sparse/missing samples, under-rate and incomplete cases
9–11	Every measured attempt counted; independent semantic validation and admission categories	Failure-latency and zero-error tests, contract corruption tests, real sanitized HTTP events
12, 14, 19	Lost-user checks, generator CPU warnings, required observations/reports, deadline and cleanup failures	Controlled incomplete evidence and setup/recovery tests; measured generator/container resources
15	Distinct stress policy; six sustained windows with reference gates	Allowed admission stress test; bad sustained window cannot hide in overall aggregate; real phase adapter smoke
16–18	Atomic evidence and retention, manifests, versions, aggregate CSV, sanitized diagnostics	51 MiB report accepted without quota, fsync failure, late marker/retention failure cannot leave a passing summary, completed/failed/active/foreign retention checks; full run reports
20, 22	Deterministic gate tests separated from measured runs; explicit qualification labels	All gate arithmetic tests and the handoff's measured result/CI-runner limitations

Qualification boundary

Reference is the complete measured local experiment. Stepped and sustained controls are checked with actual client adapters and virtual time; their capacity and 30-minute performance claims require separately selected complete runs. CI dispatch and runner qualification require a valid runner baseline. This matrix makes no production capacity, monthly availability, ANN, personalization, or training-scale claim.