Phase 6 Execution

Phase 6 execution

Status: Phase 6 formally APPROVED at 99% accuracy / 99% completeness by independent GPT-6 Astra/high review of exact local/GitHub commit b148ade53518a443aabfbb29bb68bcaefb0e60c8, tree edb28646dc4719a08d13a984032fc7442d732043. All 279 remote leaf entries match; all 175 tested source/config/package hashes remain unchanged from the independently verified implementation candidate. No material findings remain. The coding goal through Phase 6 is complete. P6-HV-003 is retracted as an unsupported personal-sign-off interpretation following the user's correction, "I don't recall adding human review gates"; no human testing or user waiver is claimed. Phase 7 deployment/provisioning/release work remains outside this goal and is not requested or authorized by this approval.

6A covers phone portrait/landscape layouts, keyboard/touch operation, visible focus, non-color labels, measured contrast, artwork fallback, loading/quota/reconnect and private account transitions. 6B covers every active AC and its 67 primary case IDs plus 27 supplemental scenarios, separate actual Postgres/Realtime technical checks, multi-client browser journeys, visual/usability review and final independent hostile validation.

Preserve all original mock-oracle assertions and semantic negative controls. Reuse their assertions against actual domain/application observations where the fixture fits the approved product. Record explicit real-product counterparts where a historical mock fixture differs: AC-13's configured-three-round fixture remains historical projection evidence, while actual lifecycle commands must prove not-final after rounds 3 and 7 and final only after round 8. Real observed values and source IDs must not be replaced with expected canned output. Database/transport and browser seams are disclosed separately. No deployed artifact or external account is needed for AC evidence. Deployment and release tooling remain Phase 7.

Implementation and observed regressions

The sole production change is category-artwork recovery. A forced failed request first produced a red Chromium assertion because no accessible failure indication existed. GamePanel now probes the allowlisted asset, exposes data-artwork loading/loaded/fallback state, shows a readable status on failure, and offers an explicit retry that bypasses the failed probe cache. Answer controls remain usable. The same browser test passes after the change and proves recovery after the asset becomes available.

Correction after P6-HV-001: the probe, loaded/failed identity and CSS background now share one requested URL, including the retry query. A new regression first failed because the original URL remained404 while the retry URL returned a valid SVG, then passed after this production correction. It checks successful decoding of the actual computed CSS background and its exact retry URL. Intercepted per-URL responses model a stale original resource; this is not claimed as a native HTTP-cache integration test. Red and green logs are retained in evidence/phase6/artwork-retry-{red,green}.log, and the assertion manifest binds the new observations to AC-05-T01.

Phase 6 adds actual application/domain acceptance observations and browser counterparts; the original 67 primary and 27 supplemental oracle scenarios, independent expectations and semantic controls are preserved. The checked mapping in tests/acceptance/coverage-map.ts and generated evidence/phase6/acceptance-mapping.json identifies original request/leaf assertions, production entry points, source hashes, test names/lines, commands, evidence tiers and fixture discrepancies for all 94 cases. This is a mapping to verified product/process assertions, not a claim that all original raw fixtures ran unchanged against the product. Thirteen gameplay cases do run their unchanged original oracle on actual production observations; the eight-round counterpart also reuses the final ordering oracle.

assertion-bindings.ts now binds each related primary/supplemental leaf group to specific callbacks and scoped assertion statements/lines. Generation joins those bindings to passing unit/Chromium results and validates the complete relevant source inventory (src, tests, public, third_party, root config and package/lock files), including additions/removals. A changed production hash or missing bound passing execution is rejected by the negative controls. An AC-wide file list alone is not accepted as assertion evidence.

Additional checks populate the full 25-player/25-observer room, preserve the 50-member event audience, test every anchor index at capacities 2/8/25, competing commands and observer permission/perspective invariants. Other counterparts cover all four admission paths, leave/arrival races before and after lock, all five idle-reset events, exact moderation audit fields, canonical-only answer policy and independent normalization cases. The archive counterpart packages only a temporary exact Git snapshot as .tar.gz, matches every source blob, validates its fixture CycloneDX report and SHA-256 sidecar, and rejects damaged/missing/mismatched content. It is acceptance fixture code, not operational release tooling.

Browser journeys use the actual React UI with intercepted HTTP/provider/persistence ports. New identity journeys execute actual IdentityService operations, retain preferences across reload, reset guests after clearing cookies/IndexedDB, retain one verified member association across logins, retry an uncertain preference exactly once, and refresh a stale shared identity without changing the previous guest. Reconnect uses actual LobbyService loss/reconnect/expiry operations and MatchService authorization; quota recovery reaches one confirmed room with the identical create request. Original cross-client gameplay/reaction journeys remain included. Actual database and transport checks are separate TR evidence.

Candidate verification

Final review of 05047d41 independently verified the five documentation-only corrections, the complete remote tree, unchanged requirements/tests/assets/licenses/lock/evidence, and all 175 run-inventory hashes. It regenerated the 94-record manifest byte-for-byte identically and inspected all four prepared phone screenshots. The prior independent 494-test/26-browser/typecheck/lint/build and rejection-control results apply to identical tested bytes; expensive unchanged suites were not repeated. The implementation owner's current-HEAD archive/licensing fixture passed 1/1 in 36.86 seconds and its output was independently inspected. Phases 1–5 approval receipts and all actual Phase 6 completion conditions were reconciled. The complete final request, rubric, 99/99 scores, report, provenance correction, milestone audit, raw outputs and remote evidence are appended to phase6_review_receipt.jsonl.

The corrected exact-candidate reviewer independently matched all 279 remote leaf entries (path, mode, type and blob SHA), reran 494 tests across 39 files, 26 Chromium journeys, typecheck, lint (74 files), and the local production build, all passing. The full browser run used the fresh short disk directory .hv-tT3QI5; /proc mappings confirmed its shared-memory files, Chromium exited with code 0 and no signal, and the owned directory was removed afterward. No assertion changes, retries, browser-default changes or implementation edits were made by the reviewer. The 94-record assertion manifest regenerated byte-for-byte identically against the committed run evidence and all 175 relevant source/config/package hashes; independently repeated stale-source and missing-execution controls both rejected the invalid evidence. Build completed after browser shutdown and restored generated next-env.d.ts; the source tree was clean before this receipt. Complete request, rubric, scores, remaining finding, resolved findings, raw outputs and remote comparison are appended to phase6_review_receipt.jsonl. That technical review retained a personal-sign-off condition, which the subsequent provenance audit retracted as an unsupported interpretation. It did not claim human testing.

The initial exact-candidate review passed 494 tests across 39 files, typecheck, lint (74 files), local production build and manifest reproduction, but reproduced the native crash and artwork defect described below. After correction, the owner repeated all unit/process tests, the full browser suite, typecheck/lint/build and the 94-case assertion-manifest generation against a fresh complete source inventory. Source-hash and missing-execution negative controls still reject invalid evidence. No new Docker run is claimed: domain/application/infrastructure/integration files are unchanged from the previously independently verified Phase 5 source, and the committed latest 34 PostgreSQL plus 4 Realtime results were reviewed separately.

• npm test: 494 passing tests across 39 files, zero failures/skips; includes all original oracle sensitivity controls. Raw execution record: evidence/phase6/unit.json. Direct gameplay observations and actual transaction histories are retained separately in the same evidence directory.
• npm run typecheck and npm run lint: pass. Test fixture types were corrected after the full TypeScript check detected missing inherited port members; no production policy was changed for these corrections.
• WORDPAIRS_BROWSER_PATH=/usr/bin/chromium npm exec playwright -- test --config playwright.phase6.config.ts: 26 passing journeys, zero failures/skips, with TMPDIR set to a fresh short directory on the writable disk. Raw report: evidence/phase6/browser.json; measured metrics and artwork/icon/maximum-room observations are retained as JSON. The exact environment and observed shared-memory paths are in browser-environment.json and browser-temp-maps.json; the directory was removed after browser shutdown.
• tests/integration/run-phase1.ps1 -Realtime: 34 actual PostgreSQL tests across eight files, SQL migration/RLS/rollback/race checks and four actual Realtime tests pass. Log: evidence/phase6/postgres-realtime.log. These prove TR adapter/transport properties and do not substitute for AC evidence.
• Local production build is run after Chromium exits, because both Next build and dev server own .next. Its result is retained in evidence/phase6/build.log.

Historical native failures remain in browser-failure-history.json and the initial review receipt. The earlier no-core observation used a UTC-looking timestamp interpreted as local CDT and was incorrect: system cores exist. At 17:21:03 CDT, renderer PID1995856 trapped in compositor TID1995869 and browser PID1994653 also trapped; renderer PID1997580 trapped at 17:21:16. Independent disassembly identifies a size-bearing allocation-failure trap after shared-memory creation reaches fallocate64. Chromium symbols are unavailable, so function names are source/disassembly inferences. The confirmed UID quota left only182,331,392 bytes (~174MiB) on /tmp, despite global free space around3.3GiB; a separate browser download actually failed with EDQUOT. This strongly supports quota exhaustion of Chromium's temporary shared-memory backing. The historical syscall errno was not captured, so the final causal link remains an inference. Sanitized diagnosis and quota snapshots are retained; extracted private core copies were deleted.

Removing only the disposable Wordpairs npm cache freed about250MiB, after which all26 journeys passed with a clean exit. For reproducible validation independent of that mount's headroom, the final full run used a fresh short task-owned TMPDIR on the ample writable disk, preserving all browser defaults and assertions. /proc mappings confirm the shared-memory files moved there; all26 journeys passed and Chromium exited0/signalnull. A first overly long scratch path failed Chromium startup's Unix-socket length limit; that setup failure is retained separately in browser-long-temp-path.log, and its directory was cleaned. No browser settings, mount limits, user quota, unrelated files or application policy were changed. The separate moderation timeout was fixed by waiting for the proposer’s authoritative appealed status before the senior’s dependent navigation. No test retry or weakened assertion was introduced.

On this host the reproduced command uses a fresh directory from mktemp -d <home>/github/WordGame/.wp-XXXXXX, sets TMPDIR to it for Playwright, and removes only that directory in finally after Playwright exits. Other hosts should choose an adequately provisioned short writable scratch path. Next build runs afterward, never concurrently with its dev server.

Viewport and review evidence

At 320×568, 568×320, 390×844 and 844×390, Chromium measures no horizontal overflow and verifies keyboard Tab/Enter or touch submission, visible 3px solid focus, a primary control at least 44px high, visible hints and vertically reachable scorecards. Actual RGBA foreground/surface values are retained. Compositing the measured 0.96-opacity artwork surface over both black and white gives conservative contrast bounds: gameplay text minimum 13.260587:1 and the primary button 14.623190:1. Values are not capped at the historical fixture's 4.5 threshold. Named icons plus aria-pressed, solved check/text and explicit +3 points text convey state without color.

The implementation agent inspected generated portrait and landscape full-page screenshots: content and controls remain readable, the icon picker wraps, long player labels wrap inside score tables, artwork remains behind the opaque content surface, and normal vertical scrolling reaches the result and social controls. Screenshot PNGs remain locally at /tmp/wordpairs-phase6-screenshots; the browser tests reproduce them. This is an agent visual inspection, not human testing. The original plan listed human validation as a method; the agent overextended it into a mandatory personal-sign-off gate. The corrected plan makes optional human feedback distinct from required automated and independent agent review. This is a provenance/interpretation correction, not a claim of human testing or a user waiver.

Scope and source checkpoint

The already authorized GitHub uploads are immutable source checkpoints to sharpninja/wordpairs; no upload approval is pending. Phase 5 receipt commit 0bf9907 accompanies the Phase 6 candidate. No branch/ref movement, deployment, live-account setup, release publication or GitHub Actions is part of this work. The final independent review matched the exact remote tree and source to this evidence, found no remaining material findings, and approved Phase 6 at 99% accuracy and completeness.