P2-A-v3-evaluator-safety-and-no-go

[SunwardL]

Summary

• Add the evaluator-only P2-A v3 semantic Memory screening change.
• Freeze the multilingual-e5-small-q8 identity, anonymous 44-record/80-query dataset, development grid, bootstrap gates, and holdout lock rules.
• Keep production retrieval unchanged: no embedding model, vector index, or runtime dependency is added.

Changes

• Add strict v3 dataset validation for scope/lifecycle labels, category quotas, zero-overlap positives, SHA-256 manifests, normalized duplicates, and cross-split near-duplicates.
• Add evaluator-only candidates for semantic-gated RRF, lexical-veto reranking, lexical-veto with a preregistered margin, and terminology-map exploration.
• Add shared authorization/lifecycle/K/prompt-budget input boundaries and fail-closed safety accounting for scope, lifecycle, authority, unknown, forbidden, fallback, and network violations.
• Add deterministic three-repeat development-grid checks and development/holdout isolation tests.
• Add local model fallback/resource evidence tests and update the v3 OpenSpec task record.

Evaluation and evidence

• Base: upstream/develop@64b648e1 (PR chore(memory):archive-merged-retrieval-evaluation-changes #1314 merged).
• Development run: 3 candidate modes × 54 frozen configurations, 40 development queries, zero network attempts, no holdout labels or metrics emitted.
• Decision: no-go. The semantic-gated control fails abstention/safety; lexical-veto matches the lexical baseline without the preregistered gain; margin reranking regresses quality. No holdout lock is generated and P2-B remains blocked.
• Refreshed evidence: D:\learning\Review_md\codex\p2a-semantic-spike\v3-development-output.json.
• Dataset query hash: 8c56c3b665b4074a389d46f910caf4118a2b5620b91c8e8458c9afa456d5419f.

Verification

• v3/related focused tests: 9 files, 32 tests passed.
• build:kun, root npm run build, full npm run lint (0 errors), changed-file ESLint, file-lines, OpenSpec strict validation, and git diff --check passed.
• Repository-wide npm run typecheck still reports the pre-existing missing phonemizer declaration in src/main/services/local-kokoro-worker-entry.ts; no production file was changed to mask it.

Scope and follow-up

• This PR contains evaluation artifacts, tests, OpenSpec, and the development no-go record only.
• Holdout evaluation is intentionally not run because no development candidate passed every preregistered gate.
• Production embedding integration and P2-B remain blocked pending a separately approved candidate version.