Outcome scorer: CJK implementation-summary pattern coverage gap

[tznthou]

Context

Counterpart to #23 (English corpus gap). Discovered during 2026-05-06 dogfood corpus audit on 39 v=2 committed/tested sessions.

Finding

After PR fixing extractOutcome to skip save-t process reports (preventing it from capturing process meta instead of real outcomes), audit shows real implementation outcomes now reach the scorer — but stay sub-threshold (0/39 hit) due to coverage gaps in 5 categories (decision-language / impl-facts / constraints / cause-effect / validation).

Common CJK implementation-summary patterns NOT covered

Sample observed in audit (post-fix, real outcomes captured but score < 2):

Pattern	Sample	Current score
Chinese commit confirmation	Commit \6de0666` 落地`	0
Bug fix commit notation	Bug fix commit = \768d95e``	1 (decision-language only)
修復總結 / 完工總結 table	## 修復總結 | 變動 | 檔案 |	1 (偶中)
Phase milestone	## Phase 1 完成 ✅ | Step | Commit |	0

Why this matters

• 0 hit / 39 ground-truth committed-tested sessions = hook auto-harvest path effectively dead for CJK dogfood users
• Counterpart to EN corpus gap (English-language scorer pattern coverage (post-#18) #23) — same root issue (regex-based scorer can't keep up with semantic richness), different language coverage

Acceptance criteria

• 3+ corroborating reports converging on the same category gap before extending anchor patterns (per English-language scorer pattern coverage (post-#18) #23 protocol — avoid dataset bias by adding patterns prematurely)
• Patterns extended must not over-broaden (false positive on noise) — verify against existing audit corpus
• Re-run audit after extension to confirm hit rate improvement

Related

• harvester writes first user prompt as memory — 70% of stored memories are prompt echoes #18 outcome scorer ship (closed)
• split low-trust session_journal from high-trust memories; add access-based decay #21 split journal/memories + access decay (open, prerequisite for proper utility validation)
• English-language scorer pattern coverage (post-#18) #23 EN corpus pattern coverage (open, sister issue)
• PR https://github.com/tznthou/ccRecall/pull/ (this fix)