docs(memory): archive semantic evaluation v3

[SunwardL]

Summary

Close out the already-merged P2-A v3 semantic Memory evaluation from PR #1317 by syncing its delta into the canonical specification and archiving the completed OpenSpec change.

PR #1317 intentionally ended at a development no-go: no candidate was locked, holdout was not run, P2-B remains blocked, and production Memory retrieval was unchanged. The change directory remained active after merge, while the main spec still described the older v2 evaluation contract.

Changes

• Sync the v3 evaluation contract into semantic-memory-retrieval-evaluation:
  • frozen zero-lexical-overlap positive stratum and cross-split near-duplicate rejection;
  • post-fix(memory): gate weak lexical relevance #1308 lexical baseline;
  • deterministic paired-bootstrap lower-bound gates;
  • fixed similarity/margin/fusion grids;
  • evaluator-only semantic-gated, lexical-veto, and margin candidates;
  • independent zero-overlap Recall reporting.
• Resolve a pre-existing specification conflict before sync:
  • rename A low-cost lexical enhancement is a first-class candidate to A low-cost lexical enhancement is exploratory;
  • preserve terminology-map evaluation as a separately reported candidate;
  • explicitly prevent it from becoming the primary candidate or authorizing production changes by itself, matching v3 task 2.6 and the implemented evaluator contract.
• Close the two conditional delivery tasks:
  • 5.3 is complete by its fail-closed branch: development produced no eligible candidate, so no lock was generated and holdout was not run;
  • 5.5 records the final green Quality and platform checks from PR P2-A-v3-evaluator-safety-and-no-go #1317.
• Archive the change as 2026-09-15-refine-kun-semantic-memory-evaluation-v3.

Scope

• No production retrieval, ranking, runtime, renderer, Main, Manager, or HTTP code changes.
• No model or runtime dependency changes.
• No edits to frozen v1, v2, or v3 datasets or decision evidence.
• P2-B remains conditionally blocked by the v3 development no-go.
• P3 feedback evolution and PR fix(instructions): harden agent context import (safety + scoped-rule fidelity) #1319 importer hardening remain separate work.

Tests

• openspec validate refine-kun-semantic-memory-evaluation-v3 --type change --strict --no-interactive --json
• openspec validate --specs --strict --no-interactive --json
• openspec validate --archived --strict --no-interactive --json
• npm run check:file-lines
• git diff --check

All validations passed. PR #1317's Quality gate and Windows, macOS, Linux x64, Linux ARM64, and final PR gate also passed before merge.