fix(release): withhold unqualified CV review and record gates

This commit is contained in:
cesnimda
2026-10-08 18:40:07 +02:00
parent 9c91575079
commit fe045b9e01
11 changed files with 544 additions and 5 deletions
+23
View File
@@ -41,6 +41,7 @@ Reconciled against [P4-06](../phase-4/p4-06/README.md#opendeferred-issue-reconci
| Backup/restore/rollback and legacy schema cutover | Production restore/cutover evidence pending | **Separate release/schema gate**; no schema in P5. Required before any external prerequisite that actually needs one |
| Stripe activation, paid deletion/retention/provider cleanup | Configuration/policy/provider proof pending; deletion remains gated | **Separate product/backend/release work**; no new billing/trial/delete UX claims or Settings migration |
| Provider mail/AI/calendar/OAuth delivery | Partial local mocks/service tests, no universal live provider certification | **Does not block unchanged navigation/presentation**; actual send/reconnect/provider semantics changes wait for sandbox/operator proof. AI truthfulness is not guaranteed by a review dialog |
| P5-13 — AI Model Semantic Qualification / AI Review UI | **DEFERRED — BLOCKED / NOT QUALIFIED**, by Connor's 2026-10-08 instruction. Local Ollama `qwen2.5:7b` is configured; deployed identity is unverified. Frozen corpus: **32 synthetic cases, 366 required calls, 0 completed calls, 0 human reviews; critical factuality failures unmeasured**. Docker Desktop remains blocked by the AF_UNIX socket reparse-point issue; no destructive recovery attempted. Deterministic grounding/proposal protections pass; P5-13 AI review UI remains unimplemented | **Outstanding technical/product debt**; [qualification and recovery evidence](p5-13-semantic-qualification/README.md), [deterministic readiness](p5-13-readiness/README.md) and exact return conditions below. Carry into P5-14/P5-15 and subsequent acceptance reports; no semantic, UI or unconditional Phase 5 acceptance |
| Full screen-reader/touch/zoom/forced-colors coverage | P4 manual browser checks partial; incomplete axe nodes remain | **Affected-slice accessibility acceptance gates** per visual protocol; unavailable evidence must stay open, not counted as success. Public certification remains blocked |
| P5-04b-R01 command entry/focus timing | One full-run failure on unchanged command/theme/test code; trace shows keyboard focus-visible gained then lost around the dialog's entry callback. Three unchanged focused reruns passed | **Separate shell accessibility debt before P5-15**, not query-population parity. [Observation](p5-04b-readiness/README.md#unrelated-baseline-observation-p5-04b-r01); no assertion weakened or shell repair attempted |
| P5-04b-R02 CV section-expansion smoke | One full-run failure on unchanged CV editor/test, before PDF export; three unchanged focused reruns passed; cause unproven | **Existing CV readiness owner**, not Jobs integration. [Observation](p5-04b-readiness/README.md#unrelated-baseline-observation-p5-04b-r02); keep the regression and investigate before broad CV migration |
@@ -48,6 +49,28 @@ Reconciled against [P4-06](../phase-4/p4-06/README.md#opendeferred-issue-reconci
| Original P4 unnamed targeted controls / preview keyboard defect | Closed for P4-02/03 targets only | **Not a blocker**, retain tests; advanced untouched controls still need their own loaded-state proof |
| Historical TS2571 | No longer present; standalone gate passed | **Not a blocker**, retain CI type gate; do not describe as current expected failure |
## P5-13 deferred debt and return conditions — 2026-10-08
**P5-13 — AI Model Semantic Qualification / AI Review UI: DEFERRED — BLOCKED / NOT QUALIFIED.** This records Connor's explicit deferral at source checkpoint `9c9157507937b3e084a7ca1e79a2ebd77d0b0928`; it does not complete P5-13 or remove its [approved requirements](PHASE5_PLAN.md#p5-13--cv-ai-review-presentation).
Local configuration selects Ollama `qwen2.5:7b`; actual deployed provider/model/digest/settings remain unverified. Docker Desktop is blocked while starting its inference manager by an AF_UNIX socket reparse-point (`IO_REPARSE_TAG_AF_UNIX`) that cannot be removed/opened normally (Windows error 1920). The deeper filesystem cause remains unknown. No destructive recovery was attempted; the failed bounded rename and supported Model Runner disable/start attempts remain documented. Do not spend another checkpoint retrying Docker recovery or repeat recovery attempts during unrelated phases.
The frozen qualification corpus contains **32 synthetic cases**. Required calls remain **366**; completed calls **0**, human reviews **0**, and critical factuality failures **unmeasured**. Zero runs cannot establish factuality. Deterministic grounding/proposal, ownership, source freshness and revision/history protections are implemented and passing independently; [fresh release-candidate regression results](release-readiness/README.md#local-validation-and-preservation) remain separate from configured-model semantic qualification. **The P5-13 AI review UI remains unimplemented.** Retain the frozen [corpus](p5-13-semantic-qualification/corpus.json), [rubric](p5-13-semantic-qualification/RUBRIC.md), [qualification results](p5-13-semantic-qualification/qualification-results.json), [initial environment evidence](p5-13-semantic-qualification/environment-checkpoint.json) and [recovery evidence](p5-13-semantic-qualification/environment-recovery-checkpoint.json), including failed observations.
Before marking P5-13 complete:
1. Verify the actual deployed provider/model/digest/settings, generation/evaluation routing, runtime/API/sidecar image revisions and prompt-source identity; local defaults and partial health diagnostics alone do not satisfy the identity gate.
2. Establish a working isolated, production-equivalent qualification environment through a supported non-destructive recovery or a verified equivalent transport. Confirm reachability and one uncached synthetic smoke request through the application routing; preserve existing privacy/consent and authentication boundaries.
3. Run the unchanged frozen corpus and rubric: all **366 required calls**, with actual uncached generation/evaluation calls and all outputs/errors retained. Preserve the three repeats per case/probe and existing assessment lane; synthetic provider regressions do not count as model calls.
4. Complete the required human review for every output/repeat, with dated reviewer evidence. An assistant review cannot replace human acceptance.
5. Resolve critical factuality failures, false approvals and unresolved ambiguity with a bounded, reviewed correction; retain failures and rerun the frozen full corpus after a justified fix. Require usable grounded rewriting and scope/language acceptance without weakening expectations.
6. Implement and validate the P5-13 AI review UI against its retained source/target, compare/edit/accept/reject/regenerate/error/locked, keyboard, theme/locale and mobile requirements.
7. Repeat relevant regression testing, preserving deterministic AI protections, ownership/privacy, revision/history, source isolation, rendering/PDF and affected workflow/accessibility contracts.
**Authorized continuation scope:** prepare and verify the completed redesign's release to the actual production branch, **`main`**, first, subject to every production/release gate. Repository inspection identifies [the configured `main` CI/deployment workflow](../../../.gitea/workflows/ci-deploy.yml); a `main` push triggers automatic production deployment, so merge/push remains **NO-GO pending production safety proof**, including F04 ingress and backup/restore/rollback evidence. The requested `master` wording does not override the instruction to determine the actual configured branch. Only after successful release verification may a new development branch from updated `main` continue **P5-14 for currently implemented CV capabilities**: desktop/tablet/mobile editor-preview switching, section editing/navigation, Career Profile source/override clarity, save/conflict recovery, keyboard/pointer/touch, document controls, preview scroll/zoom, PDF fidelity, accessibility and localisation. This is a conditional scope clarification; the [historical P5-14 requirement](PHASE5_PLAN.md#p5-14--responsive-cv-workflow-checkpoint) for accepted 09–13 and a complete AI-review journey remains recorded. Do not fabricate an AI-review workflow, claim deferred P5-13-specific responsiveness passed, or remove the existing automated AI protection tests. Carry that missing coverage into P5-14 acceptance explicitly.
P5-15 must carry this debt and distinguish completed/verified capabilities, explicitly deferred capabilities, production-release blockers and non-blocking limitations. Full Phase 5 acceptance remains conditional/partial while mandatory retained requirements are unmet; no unconditional GO without an explicit documented authorization revising the full acceptance scope. The release/P5-14 clarification does not waive F04, release safety, human accessibility acceptance or unrelated debt. Preserve Connor's requirement to **push every individual P5-15 change/checkpoint**; outside P5-15, commit and push at each phase end. No release or deployment success is asserted by this register update.
## Product/new-requirement boundaries
Final name/logo/primary color remain OD-01–03: use current provisional values, so no shell delay. Email provider (OD-05), external jobs/calendar integrations (OD-06), monetization (OD-07), favorite/priority (OD-08), aggregate admin/privacy definitions (OD-09), intake (OD-10) and weekly goals (OD-11) are separate decisions before their future features. Do not invent LinkedIn integration or scheduled interviews.