Split Phase 5 browser evidence without rebinding the candidate

[Blazenetic]

What changed

Post-Phase-5 engineering review of source candidate ae9f7605a9cf613695c4363faa3e5250b67673a8. Evidence, verifier and documentation only. The built application is unchanged, so CI run 28 remains a valid browser artefact.

• Split the 29-criterion EV-LOCAL-BROWSER-BLOCKED blob into illustrative / offline-shell / interrupt passes plus a leftover blocked record.
• Promote only criteria run 28 actually established (hero loop, named controls, table captions, reduced motion, 24px targets, Chromium/Firefox cached shell, fixture-load interrupt, pause-before-selection unit test).
• Leave live DistilGPT2, WebGPU-denied, visual inspection, complete tab order, current screenshots and related criteria blocked or not-run.
• Mechanical freeze: candidate must be an ancestor of HEAD; apps/, packages/, fixtures/, tests/ and the lockfile may not drift.
• Reject passed records whose limitation still says the work remains blocked.
• Assert hero token 3223 / " dark" in pnpm trace:verify:live.
• Structured review: docs/implementation/06-post-phase-5-review.md.

Scientific-integrity check

• Live, demo, derived, projected, probed and interventional outputs are clearly labelled.
• No new copy anthropomorphises model behaviour or overstates causality.
• Any new metric has a documented method, units and compatibility boundary.
• Fixture-worker Playwright journeys are not treated as DistilGPT2 measurements.
• TMO-FUNC-001 is passed as the illustrative teaching fixture opening immediately; the criterion text still says “verified trace” — owner decision whether to reword or re-block.

Verification

Observed on Node 24.19.0 / pnpm 11.19.0:

• pnpm phase5:verify — 78 criteria, 22 launch blockers, decision NOT READY
• pnpm test — 82 passed (was 78; four new verifier regressions)
• pnpm trace:verify:live — token 3223 (" dark") from 50,257 logits
• pnpm lint, pnpm typecheck, pnpm fixtures:check, pnpm payload:verify, pnpm build
• pnpm phase5:budgets — local gzip of unchanged app JS was 108,783 vs candidate-bound 108,752; CSS 6,901 and worker 922,478 matched. Report not rewritten. GitHub CI is the budget authority.
• pnpm e2e — not re-run; would invalidate the “same artefact as run 28” claim if app files had changed, and they did not.

Result counts: 51 passed / 0 failed / 16 blocked / 11 not-run. Was 38 / 0 / 29 / 11.

Evidence

• Generated summary: release-evidence/summary.md
• Manifest split: EV-CI-PLAYWRIGHT-ILLUSTRATIVE, EV-CI-PLAYWRIGHT-OFFLINE-SHELL, EV-CI-PLAYWRIGHT-INTERRUPT, leftover EV-LOCAL-BROWSER-BLOCKED
• Browser artefact: run 28, Playwright 1.62.1, 33 passed / 6 skipped
• Review: docs/implementation/06-post-phase-5-review.md

Known limits / next slice

Do not tag this candidate.

Evidence still needed on this SHA: hero screenshots, deployed-origin smoke, named screen-reader, physical Android/Safari, RUN_LIVE_MODEL=1, learner study first-branch-v1.

Next source candidate (rebinding required): 1.1 PRNG replay cursor, compact cursor chaining, live logit hashing, cache-first navigation, skip-target outline, unique Force/Suppress names, real tab-order journey.

Hidden states, attention, probes and projections stay unavailable.