fix(cognition): whole-word merged in episode classifier (recall quality / distillation fact-yield)

[rysweet]

Summary

Perpetual-cognition improvement on the recall-quality / distillation-fact-yield axis.

The episode-ingestion classifier (memory_consolidation::classifier::classify) decides — at every store_episode intake site — whether an event is operational noise (DROP), low-value bookkeeping (DOWN-SCOPE), or a meaningful episodic worth full-importance storage (STORE). Distillation later mines the STORED, high-importance episodes into semantic facts, so a misclassification here propagates straight into recall top-k.

The defect

Rule 3's durable-completion signal scored a single token with a raw substring content_lc.contains("merged"). That fired inside the git-status vocabulary unmerged / submerged — which name the opposite of a completed merge (an outstanding, un-completed one) — promoting benign/negative bookkeeping to a durable ActionCompleted episode at 0.7 importance. Distillation would then mine that phantom completion into a phantom "action completed" fact, polluting recall top-k and dragging recall precision down.

The fix

merged now matches at word boundaries via a new contains_word helper, aligning this seam with the word-boundary policy already adopted by the same classifier's failure-signal pass (#4080) and the knowledge-pack relevance scorer (#4241). unmerged/submerged down-scope to the operational tier; a genuine whole-word merged still classifies as a durable completion, so no real completion changes disposition. The multi-word phrases (opened pr, pull request) cannot embed in a single word, so they intentionally remain substring checks.

Merge-ready evidence

Criterion 1 — qa-team scenarios (validated + run):

• New outside-in scenario tests/gadugi/episode-classifier-whole-word-completion.yaml drives the contract over the real classifier (headline asserts unmerged_paths_are_not_a_durable_completion ... ok, whole_word_merged_is_still_a_durable_completion ... ok, and the unchanged meaningful_completed_action_is_stored, plus test result: ok + exit 0).
• The gadugi-test binary is not installed in this engineering environment, so gadugi-test validate/run could not be executed here. The scenario was validated structurally (YAML parses; schema-parity with the sibling distill-fact-yield-informative-content.yaml) and its underlying command cargo test --locked --lib memory_consolidation::classifier passes (20 passed; 0 failed).

Criterion 2 — docs / changed-surface justification: No user-facing surface changed. classify signature/contract is unchanged; this refines a private helper and one internal branch only. No CLI/operator/README surface touched → internal-only, no doc update required.

Criterion 3 — quality-audit (SEEK→VALIDATE→FIX): 3 cycles. Cycle 1: found the unmerged/submerged substring false-positive → added contains_word + tests. Cycle 2: verified no other single-token durable-completion signal has the same class (multi-word phrases are embed-safe; handoff has no embedding words and would regress handoffs under exact-match, so left as substring; recipe only routes within the failure branch). Cycle 3: caught + fixed a doc-comment inaccuracy (importance band 0.9→0.7). Final cycle clean: zero critical/high; zero medium correctness/security findings.

Criterion 4 — CI green (local gates mirror CI):

• cargo fmt --all -- --check ✅ (pre-commit + pre-push)
• cargo clippy --release --no-deps -- -D warnings ✅ (pre-commit)
• cargo clippy --all-targets --all-features --locked -- -D warnings ✅ (pre-push)
• Race-subset release tests cognitive_memory bootstrap memory_ipc memory_consolidation ✅ (pre-push)
• cargo test --locked --lib memory_consolidation → 148 passed; 0 failed.

Criterion 6 — focused diff: 3 files (classifier.rs production, classifier_tests.rs tests, one new gadugi scenario). No unrelated edits.

Tests added

• unmerged_paths_are_not_a_durable_completion — unmerged/submerged do NOT store as ActionCompleted; down-scope to the operational tier.
• whole_word_merged_is_still_a_durable_completion — a genuine whole-word merged still classifies as ActionCompleted (no regression of the real signal).

Standing perpetual goal — durable code improvement, not a snapshot doc.

Co-authored-by: Copilot 223556219+Copilot@users.noreply.github.com

[github-actions]

📊 Coverage Summary

Generated by cargo llvm-cov --workspace --summary-only (nightly, excluding test files)

Module	Lines	Covered	Coverage
Total	184497	153960	83.4%

_{Coverage data from CI run. Test files matching tests?/ are excluded from line counts.}