wiki-audit: knowledge pipeline hardening — gate bypass window, checkable dedup evidence, harvest caps

[choiyounggi]

Finding (wiki-audit 2026-08-05, axis: harvest/flush pipeline)

Fixed on branch fix/harvest-dedup-processed: processed-store dedupe in harvest.js, quoted/$HOME --body-file parsing in pre-flush-pr-gate.sh, SKILL literal-path guidance, retire-dropped-candidates clarification (21 bats tests).

Residual hardening candidates (plans/wiki-audit/findings/pipeline.md):

1. Gate bypass window: flush detection is command-marker-only (--head knowledge/, label, INGEST_REPORT). A gh pr create --body inline with no label and head inferred from the current branch engages nothing. Tightening options: also match --title "knowledge:"; have the flush prompt forbid --body.
2. Existing-layer check is free prose: the gate proves the section exists, not that wiki pages were read. Option: require the report to list page ids read, machine-checkable against wiki/**.
3. No per-session harvest cap: the instruction says 0–3 insights/session but harvest.js enforces none; a runaway session could queue dozens and burn a headless auto-flush run.
4. Empty per-session queue files accumulate (~35 observed) — cosmetic cleanup in harvest or flush.

[choiyounggi]

Real-world evidence observed right after merging #41: 15 open knowledge/* PRs (#17–#40) with heavily overlapping content — "client-side rate limiting" appears in ≥5 PR titles (#22 #24 #29 #32 #34), "call-site enumeration" in ≥6 (#22 #23 #24 #25 #29 #32).

Two mechanisms produced this:

1. Pre-fix(knowledge): harvest processed-store dedupe + flush-gate hardening (wiki-audit) #41 re-queuing — every Stop re-parsed the transcript and re-queued flushed hashes, so successive auto-flushes kept re-ingesting the same insights. Fixed by fix(knowledge): harvest processed-store dedupe + flush-gate hardening (wiki-audit) #41 (processed-store dedupe) going forward, but the existing PR pile predates it.
2. No cross-PR dedup (still open) — flush step 2b checks the merged wiki (origin/main) for duplicates, never the pending sibling knowledge/* branches. Two sessions emitting similar insights (different hashes) each open a PR that duplicates the other. Suggested hardening: before ingest, also diff against open knowledge/* heads (gh pr list --state open --search 'head:knowledge/') and either amend the existing PR's branch or drop the candidate as pending-duplicate.

Near-term cleanup: the 15 open PRs need a consolidation pass (same procedure as the earlier #6–#13 reconcile) — most are supersets/subsets of one another.