Eve library

Release evidence run — 2026-09-12

Guide

Revision 1

On this page

Evidence for 182.C.40.09: run and archive all applicable targeted lint, typecheck, unit, integration, contract, build, Playwright/mobile, accessibility, performance, provider-canary, evaluator, and release-gate evidence.

Twelve categories. Each was run in this session unless the row says otherwise, and each says what "applicable" means for it.

# category result how
1 targeted lint 0 errors, 42 warnings eslint over all 504 human-video-*.ts, in three batches
2 typecheck 0 errors all 504 files in ONE program against tsconfig.typecheck.json, with a deliberately broken control file that reported
3 unit 5442 BFF (262 files), 215 web (11), 128 mobile (8) vitest, jest
4 integration included above the upload vertical and the runtime over real Fastify instances with a filesystem sink
5 contract 40 tests + 2 guards SDK byte-identical drift guard, consumer contracts, verify-client-api-calls.mjs (13 URL shapes across 7 hand-written clients), verify-mobile-review-automation.mjs
6 build passes, 11.7 s node scripts/build.mjs in apps/oshun/bff
7 Playwright 11 passed / 1 skipped web, 5 passed admin against real stacks; servers torn down
8 mobile 128 unit + 2 verifiers the Maestro flow itself needs a device and is in the limitation table
9 accessibility 0 violations axe inside both e2e suites, at a stricter threshold than the suite default on the Studio
10 performance 46 tests capacity model, soak, benchmark harness
11 provider canary archived evidence/2026-09-11-wan-canary/, $0.50 measured against a $0.20 prediction
12 release gate 50 tests final-bytes gate, promotion gates, voice-use gate

What running it as one program found#

Three real lint errors that per-file runs had missed. All three were no-useless-assignment — a value assigned and never read — and none had ever appeared in a per-item lint, because a per-item lint only ever saw the file that item touched:

  • human-video-metadata-strip.ts:141,148 — cursor assigned immediately before a break, twice. Harmless to behaviour and removed, because an assignment that suggests a position somebody might later depend on is worth not having.
  • human-video-take-comparison.ts:179 — let within = 0, where every path below either returns or assigns. Now declared without a value: 0 is a real position inside a word, so a default surviving a missed branch would read as "at the very start of it" rather than as a bug.

The typecheck over all 504 files in one program found nothing — which is itself the result worth recording, because the same exercise on 2026-09-11 found 32 real errors that had accumulated unread across 18 files.

Categories where "applicable" is doing work#

  • Playwright/mobile: the Maestro flow needs a device. What runs here is the verifier that keeps its test ids and the screen in step, which is what mobile-parity.md records as the reachable half.
  • Performance: the capacity model and soak tests are models and simulations, not a load test against a running provider. There is no deployed lane to load.
  • Provider canary: not re-run. One billed render is archived from 2026-09-11 and re-running it would spend money to re-learn a measured fact.
  • Accessibility: the automated half. The manual screen-reader, zoom and contrast passes 182.C.36.08 asks for are not automatable and that item stays open.

What this does not establish#

Nothing here is a demonstration that the phase's product works end to end — that is 182.C.40.01–.08, and they need jobs, candidates and release adapters that do not exist. This says the code that IS written builds, checks, lints and passes its own tests, and that the guards which keep it honest fire.