fix(knowledge): close the learning feedback loop

[zaxbysauce]

Closes #1715

Summary

The knowledge feedback loop was an open circuit — outcome signals were recorded but nothing acted on them. Three gaps fixed (G1/G2/G3 from the gap catalog):

• G1 — phase-failure outcomes now record. phase-complete previously called updateRetrievalOutcome(dir, name, true) with a hardcoded true at a point in the flow before success was determined. The call now runs after the real outcome is finalized and passes that value, so a failed phase (policy=enforce + missing required agents) emits 'failure' outcome events and failed_after_shown_count can finally increment.
• G2 — confidence-floor action. Both feedback bridges (applySkillUsageFeedback, applyKnowledgeVerdictFeedback) bumped entry.confidence but nothing acted on low confidence — the bumps dead-ended at the floor. Now bumpKnowledgeConfidenceBatch runs a post-bump sweep that demotes (confidence_floor_demoted flag → strips retrieval statusBoost) or quarantines per config entries that are floor-clamped AND have a net-negative outcome signal over ≥3 outcome events. Reversible by default.
• G3 — contradiction signals unified. contradicted_count (incremented only via knowledge_receipt) and the curator's flag_contradiction (tag-only) were disconnected. The curator now emits contradicted events post-transaction (with agent: 'curator' for audit), and a new maybeQuarantineOnContradiction auto-quarantines entries whose in-window contradicted count crosses the threshold (default 3 in 30d). Counts come from the RAW event log, not the mtime-keyed rollup cache, to avoid stale-count risk.

The genuinely-closed sub-loop was previously only the violation-count → escalator → enforce-block path. This change closes the confidence/contradiction/phase-failure halves.

Invariant audit

• 1 (plugin init): not touched — no plugin-registration changes.
• 2 (runtime portability): not touched — pure TS, no runtime-specific code.
• 3 (subprocesses): not touched.
• 4 (.swarm containment): not touched — new artifacts (contradicted events, confidence_floor_demoted flag) write only to existing .swarm/ knowledge + events files; no new locations.
• 5 (plan durability): not touched.
• 6 (test_runner safety): not touched.
• 7 (test writing): touched — 4 new/updated test files co-located with source per convention; each runs in its own process to avoid the cross-file contamination tracked in coverage job's full-suite single-process run shows ~189 files (16%) with test failures — likely systemic mock/state contamination, not independent bugs #1712.
• 8 (session state): not touched.
• 9 (guardrails/retry): not touched — fail-open semantics preserved (bumpKnowledgeConfidenceBatch and the new floor-action are best-effort, never throw).
• 10 (chat/system msg): not touched — the G2 statusBoost gate is read-only on retrieval scoring.
• 11 (tool registration): not touched.
• 12 (release/cache): touched — release fragment added at docs/releases/pending/1715-close-knowledge-feedback-loop.md; no version files edited.

Test plan

Validation commands run + results:

$ bun run typecheck                    # tsc --noEmit
(empty — clean)

$ bun run lint                         # biome lint .
Checked 2313 files. No fixes applied.

# Each affected test file in its own process (per repo convention; the
# combined single-process run has pre-existing cross-file contamination
# tracked in #1712 — see proof below).
$ bun --smol test tests/unit/config/knowledge-config.test.ts            # 43 pass / 0 fail
$ bun --smol test tests/unit/hooks/knowledge-reader.test.ts             # 20 pass / 0 fail
$ bun --smol test tests/unit/hooks/knowledge-events.test.ts             # 24 pass / 0 fail
$ bun --smol test tests/unit/hooks/knowledge-store.test.ts              # 39 pass / 0 fail
$ bun --smol test tests/unit/hooks/knowledge-verdict-feedback.test.ts   # 10 pass / 0 fail
$ bun --smol test tests/unit/hooks/curator.test.ts                      # 118 pass / 0 fail
$ bun --smol test tests/unit/hooks/knowledge-escalator.test.ts          # 5 pass / 0 fail
$ bun --smol test tests/unit/hooks/knowledge-floor-action.test.ts       # 6 pass / 0 fail   (NEW)
$ bun --smol test tests/unit/hooks/knowledge-contradiction.test.ts      # 8 pass / 0 fail   (NEW)
$ bun --smol test tests/unit/hooks/search-knowledge.test.ts             # 14 pass / 0 fail
$ bun --smol test tests/unit/tools/phase-complete-outcome-regression.test.ts  # 3 pass / 0 fail  (NEW)
                                                                        # = 290 pass / 0 fail

Pre-existing-failure proof (combined run is NOT a regression): the combined single-process run produces 24 failures — but I verified this is identical on clean origin/main (worktree at HEAD c5eb59c, before any of these edits): same 24 failures, same pass count. This is the documented cross-file mock/state contamination tracked in #1712, not caused by this change.

Independent review: GLM-5.2 reviewer (fresh context) returned NEEDS_REVISION with 5 findings (2 HIGH broken config fixtures + missing tests; 1 MED hive quarantine no-op; 2 LOW). All addressed. GLM-5.2 critic (separate fresh context) returned APPROVE with 2 optional LOW improvements, both applied (floor-action console.warn, behavioral retrieval-ranking test).

New test coverage:

• G2: floor-demote on net-negative signal, min-outcomes gate, net-positive skip, recovery clears stale flag, none=legacy, idempotency, retrieval down-ranking.
• G3: curator emits contradicted event, tag preserved (back-compat), threshold quarantine, below-threshold skip, idempotency on already-quarantined, window respected, curator→quarantine end-to-end, tag_only=legacy.
• G1: source-level regression guard (no literal true, passes success var, old false-premise comment gone). Companion behavioral test (false → 'failure' event) already exists at knowledge-reader.test.ts Test 12.

How to use

New config knobs (all optional, conservative defaults):

• knowledge.confidence_floor_action: 'none' | 'demote' | 'quarantine' (default 'demote').
• knowledge.confidence_floor_min_outcomes: default 3.
• knowledge.confidence_floor_signal_threshold: default 0 (net-negative).
• knowledge.contradiction_threshold_action: 'tag_only' | 'quarantine' (default 'quarantine').
• knowledge.contradiction_quarantine_threshold: default 3.
• knowledge.contradiction_quarantine_window_days: default 30.

For the legacy dead-end behavior: confidence_floor_action: 'none' + contradiction_threshold_action: 'tag_only'.

Migration

No migration required. New config fields default to conservative behavior. Existing entries without confidence_floor_demoted read as undefined → not-demoted (legacy). Hive floor-entries get the flag (honored by retrieval) but not quarantine (the quarantine mechanism reads only the swarm file — documented in the release fragment).

[github-actions]

Drift check report

✅ No drift detected across skills, tools, commands, and agents.

[zaxbysauce]

🤖 Multi-Stage PR Review

Pipeline: Friday-35B (Qwen3.6-35B-A3B) (context pack) → Mellum2-12B (explore → candidates) → GLM-5-turbo (challenge + blind-spot)
Commit reviewed: 8c8d90808f31

---

🔍 PR Intent

Reconstructed obligation list from PR text, issue #1715, commit messages, changed tests, changed docs, and changed interfaces.

• O-001 phase-complete must call updateRetrievalOutcome with the real success variable, not a hardcoded true (G1).
• O-002 When a confidence-bumped entry is clamped to the floor (0.1) with a net-negative outcome signal and ≥3 outcome events, it must be demoted (default) or quarantined per config (G2).
• O-003 Curator flag_contradiction actions must emit contradicted events post-transaction to unify the two previously-disconnected contradiction signals (G3).
• O-004 When an entry's in-window contradicted count crosses the configured threshold, it must be auto-quarantined (G3).
• O-005 New config knobs must have correct Zod schema definitions with proper defaults.
• O-006 The confidence_floor_demoted flag must suppress statusBoost in retrieval ranking.
• O-007 Recovery above the floor must clear a stale confidence_floor_demoted flag.
• O-008 Backward compatibility: entries without confidence_floor_demoted treated as not-demoted; none/tag_only config options preserve legacy dead-end behavior.

---

📦 Implementation Summary

Three independent feedback-loop gaps are closed. G1: The updateRetrievalOutcome call in phase-complete.ts is moved after success is finalized and passes that variable instead of true; a source-level regression test guards against re-introduction. G2: bumpKnowledgeConfidenceBatch now returns the set of touched entries from applyConfidenceDeltas and runs a post-bump applyConfidenceFloorAction sweep that reads event rollups once, checks signal/evidence thresholds, and persists confidence_floor_demoted flags (and optionally quarantines) via separate transactKnowledge calls for swarm and hive tiers. Retrieval scoring in search-knowledge.ts suppresses statusBoost when the flag is set. G3: applyCuratorKnowledgeUpdates collects flag_contradiction actions during its transaction, then emits contradicted receipt events post-transaction (avoiding lock deadlock) and calls maybeQuarantineOnContradiction which counts raw events and auto-quarantines at the threshold. Six new config fields are added to the Zod schema with conservative defaults. New type ConfidenceFloorOptions threads the config through both feedback bridges.

---

✅ / ⚠️ / ❌ Intended vs Actual

Obligation	Status	Evidence (file:line)
O-001	SUPPORTED	src/tools/phase-complete.ts:1383 — passes success variable; old literal-true call removed
O-002	SUPPORTED	src/hooks/knowledge-store.ts:906–1089 — applyConfidenceFloorAction with signal/evidence gates, demote/quarantine/none paths
O-003	SUPPORTED	src/hooks/curator.ts:1718–1745 — emits contradicted events with agent:'curator' post-transaction
O-004	SUPPORTED	src/hooks/knowledge-escalator.ts:280–352 — maybeQuarantineOnContradiction with raw-event counting, idempotency guard
O-005	SUPPORTED	src/config/schema.ts:1102–1131 — six new fields with correct Zod types and defaults matching docs
O-006	SUPPORTED	src/hooks/search-knowledge.ts:487–491 — !entry.confidence_floor_demoted gate on statusBoost
O-007	SUPPORTED	src/hooks/knowledge-store.ts:943–944,1014 — recovered array clears stale flag via entry.confidence_floor_demoted = false
O-008	SUPPORTED	undefined is falsy → not-demoted (legacy); floorAction:'none' early-returns; tag_only skips quarantine call

---

🚨 Confirmed Findings

None found.

---

🔬 Unverified but Plausible Risks

None found.

---

🧪 Test / Coverage Gaps

None found. All three gaps (G1 source guard + behavioral companion, G2 demote/quarantine/recovery/idempotency/none, G3 event emission/tag-backcompat/threshold/below-threshold/idempotency/window/end-to-end/tag_only) are covered. The retrieval-ranking test confirms the demoted entry actually sinks.

---

📋 Shipped-vs-Claimed Gaps

None found. 290 tests pass (all in isolated processes per repo convention). New config defaults match the schema, the release fragment, and the PR description exactly.

---

🔁 Validation provenance

• Confirmed findings kept: 0 (the reviewer-confirmed list was empty).
• Confirmed findings dropped: N/A.
• Blind-spot findings added: 0. I traced every + line across all seven changed source files, verified cross-file type contracts (ConfidenceFloorOptions wiring), checked lock ordering (sequential, no deadlock), validated the touched-array propagation from applyConfidenceDeltas → applyConfidenceFloorAction, confirmed hive-tier quarantine is intentionally omitted (documented caveat), and verified the G1 guard condition is a semantic tightening consistent with the function's purpose. No defects found.

---

📝 Merge Recommendation

APPROVE

Clean, well-structured closure of three documented feedback-loop gaps with conservative defaults, proper fail-open semantics, correct cross-file wiring, and thorough test coverage.

Check	Result
No CRITICAL findings	✅
No unresolved STEALTH_CHANGE	✅
No UNSUPPORTED obligations	✅
Test coverage adequate	✅
No hardcoded secrets	✅
All async errors handled	✅
Input validation present	✅
No broken agent role boundaries	✅
Prompt format contracts intact	✅
Lockfile consistent	✅

---

_{🔒 Reviewed by a multi-stage local-first funnel (architect context pack → explorer candidates → critic challenge/author) for high recall with low false-positive noise. Findings are advisory — verify before acting.}