Wrap-up flow leakage: Chinese verb form '存檔完成' bypasses PROCESS_REPORT_RES anchor

[tznthou]

Context

After v0.3.0 P1 ship (PR #26 / closes #21) split persistence into trust-graded session_journal (low-trust) → manual ccmem promote → memories (high-trust), the KNOWLEDGE_THRESHOLD persistence gate was removed. Score 0/1 entries now reach session_journal as designed. To compensate, outcome-scorer.ts keeps a hard-floor isProcessReport() filter (defense in depth) that rejects session wrap-up process reports before they enter the journal.

Finding

24h after v0.3.0 daemon upgrade (5/6 evening → 5/7 09:18), 11 entries reached session_journal. 3 of them (id=2, id=6, id=8 — 27% of intake) are session wrap-up process reports that bypassed isProcessReport().

Root cause

src/core/outcome-scorer.ts:28-30:

const PROCESS_REPORT_RES: ReadonlyArray<RegExp> = [
  /^[#\s💾🟢✅_*-]*\/?save[\s-]?t(?:\s*[:：]\s*|\s+)(?:完成|流程完整|完整|done|finished|completed?)/iu,
]

Anchor literally requires the slash-command literal at start. But the wrap-up flow's assistant-side output in CJK opens with the Chinese verb form 存檔完成 instead of the slash form the anchor expects. The regex misses it entirely.

Evidence — 3 corroborating reports (24h)

journal id	created_at	content opening	score	reasons
2	2026-05-06 07:20:38	存檔完成 — 雙寫同步、diff 為空。	0	[]
6	2026-05-07 04:34:27	存檔完成，可以安全 \/compact` 或結束 session。`	0	[]
8	2026-05-07 04:48:29	存檔完成。\n\n## 更新摘要\n\n| 檔案 | 動作 |	0	[]

All 3 share the pattern: short verb-form opener (存檔完成) + optional dash/comma clause + table or bullet list. score=0 (no signal category hit), so they only made it through because the persistence gate is gone — which is the intended trust-grade design, but isProcessReport() was supposed to short-circuit them.

Manual cleanup tested: ccmem reject 2 6 8 → all marked rejected, expires 2026-05-14 (sweep clears).

Why this matters

• 27% noise-rate in journal directly hurts journalPendingCount signal — 5/20 14d observation gate (COUNT(*) >= 50) loses signal-to-noise resolution.
• Distinct from Outcome scorer: CJK implementation-summary pattern coverage gap #25 (CJK implementation-summary scorer pattern coverage) — that issue is about real outcomes scoring sub-threshold. This is about process meta bypassing the hard-floor filter.
• Distinct from the original 5/6 morning v0.2.7 fix (which added PROCESS_REPORT_RES to anchor the slash-command literal) — fix targeted assistant text starting with the slash form; missed the Chinese verb form.

Proposed fix

Extend PROCESS_REPORT_RES with verb-form anchors, narrow enough to avoid false-positive on real text containing 存檔完成 mid-sentence:

const PROCESS_REPORT_RES: ReadonlyArray<RegExp> = [
  /^[#\s💾🟢✅_*-]*\/?save[\s-]?t(?:\s*[:：]\s*|\s+)(?:完成|流程完整|完整|done|finished|completed?)/iu,
  /^[#\s💾🟢✅_*-]*存檔(?:完成|完畢)\s*[—,，.。:：!！]/u,
]

Risks:

• False-positive on legit 存檔完成。然後我修了 X bug opener — accept since real impl-facts content rarely opens with this phrase
• Pattern over-broadening — keep narrow with required punctuation/em-dash after 存檔完成

Acceptance criteria

• Add regex + 3 ground-truth should_reject cases (id=2/6/8 content) to tests/outcome-scorer.test.ts
• Add 1 false-positive guard test: long-form text containing 存檔完成 mid-sentence should NOT match
• Verify journalPendingCount post-fix on next 24h dogfood run shows wrap-up verb-form pattern at 0
• No regression on existing slash-command literal anchor

Related

• split low-trust session_journal from high-trust memories; add access-based decay #21 (closed by PR P1 (#21): split low-trust session_journal from high-trust memories #26) — v0.3.0 trust split that removed persistence gate
• Outcome scorer: CJK implementation-summary pattern coverage gap #25 (open) — CJK implementation-summary scorer pattern coverage (different layer)
• English-language scorer pattern coverage (post-#18) #23 (open) — English scorer pattern coverage (counterpart for EN corpus)

3+ corroborating reports converging on 存檔完成 opener already met within 24h — fix can proceed without waiting for additional samples.