feat(mycelium): add distill subcommand + harden pr-retro capture quality

[howie]

What & Why

Adds a Hermes-style knowledge-distillation loop on top of the existing mycelium
lessons table, modeled on NousResearch/hermes-agent's Skills System (procedural memory +
periodic nudges + autonomous skill-creation prep). /pr-retro already captures per-PR
lessons, but nothing periodically reviews accumulated lessons to surface recurring patterns —
this fills that gap (the synthesis/skill-drafting half lands in ainization-skill as the
knowledge-distill skill).

Part 0 — pr-retro capture hardening (upstream signal quality)

The distill scoring relies on confidence and recurrence signals that /pr-retro was
flattening:

• --confidence no longer hardcoded to 7 — differentiate by source (user-stated 8-9,
  cross-model 8, inferred 5-6; +1 on recurrence)
• Q4 lessons unlock from a fixed "3 points" to 0-5 (quality over padding)
• lessons add --skill records the subject skill, not the producer (pr-retrospective)
• --key slug gains a domain-prefix convention so recurrence/dedup converge across PRs

Part A — mycelium distill subcommand

• distill_service.py: harvest (read-only time window) → cluster (deterministic
  token-Jaccard + type + key-prefix; pluggable _similarity for when embeddings land) →
  score (≥3 members, ≥2 distinct retro_pr, avg_conf ≥7, procedural type, new-evidence
  since watermark) → DigestReport JSON
• models.py: DistilledCluster / SkillCandidate / DigestReport
• cli.py: distill run (read-only + watermark) and distill promote-tiers (wires the
  previously unscheduled tier promotion)
• config.py: DISTILL_DIR / DISTILL_STATE_PATH
• watermark state.json makes runs idempotent — periodic nudge only fires on new accrual

Tests / verification

• 20 new tests (clustering, score thresholds DT, watermark idempotency); full mycelium suite
  green (389)
• ruff / mypy / pylint / bandit / lint-skill-bash / markdownlint all green
• Real-data dry-run over 233 lessons surfaced 4 coherent cross-PR candidates

🤖 Generated with Claude Code

[howie]

Final Aggregated Review — PR #173

Mode

group-review (3/3 voices active: Claude 4-subagent / agy / Codex)

Voice verdicts

• Claude (code-reviewer + silent-failure-hunter + pr-test-analyzer + comment-analyzer): NEEDS_CHANGES
• agy / Gemini: NEEDS_CHANGES (2 Critical refuted on verification, 1 consensus Important)
• Codex: LGTM (diff inspected, no substantive findings; test exec blocked by sandbox temp dir)

Consensus Important (must fix)

1. Prefix union-find over-clusters (Claude code-reviewer + agy) — distill_service.py cluster():
   same_prefix unconditionally unions any two same-type lessons sharing a key prefix, regardless
   of token overlap. With union-find transitivity, all bash-* pattern lessons collapse into one
   grab-bag mega-cluster, diluting candidates (contradicts the PR's own "湊數稀釋 cluster" goal).
   Fix: make prefix lower the similarity bar, not override it — union iff
   _similarity >= threshold OR (same_prefix AND _similarity >= PREFIX_FLOOR) (PREFIX_FLOOR=0.15).
   Update DT-002 to assert corrected behavior + add a zero-overlap-same-prefix → no-merge test.

Single-voice, verified real (must fix)

2. [test gap, raised Critical] Positive watermark resurface path untested (Claude pr-test-analyzer)
   — verified by inspection: every "candidate present" test uses wm=None; the periodic-nudge
   positive branch (wm != None AND a member newer than wm → candidate surfaces) has zero coverage.
   An inverted > would pass the whole suite. Fix: add a resurface test (watermark older than
   members → 1 candidate; advance past → 0; insert fresh post-watermark member → resurfaces).

3. Silent data loss on parse failures (Claude silent-failure-hunter; agy raised the Z-suffix
   variant) — _parse_ts returns None and harvest folds unparseable-ts rows into the same
   "skip" branch as out-of-window rows; load_watermark swallows corruption → silent re-flood +
   overwrite. Fix: count + surface dropped-unparseable-ts (CLI warn + DigestReport field);
   distinguish corrupt watermark from missing (warn before treating as first run).

4. Silent harvest truncation at 2000 (Claude code-reviewer + silent-failure-hunter) — warn when
   len(rows) == _HARVEST_SCAN_LIMIT so window truncation is observable.

5. SKILL.md confidence rubric contradicts hardcoded --source inferred (Claude comment-analyzer)
   — the new by-source confidence guidance is impossible to honor while --source stays hardcoded.
   Fix: parametrize --source in the lessons-add template alongside --confidence.

6. /knowledge-distill forward reference (Claude comment-analyzer) — referenced as the downstream
   consumer but lives in the separate ainization-skill repo. Fix: clarify cross-repo / planned.

Disputed → refuted with evidence (drop)

• agy [Critical] from datetime import UTC breaks Python < 3.11 → REFUTED: pyproject.toml
requires-python = ">=3.11"; existing modules already import UTC; CI runs 3.12. The Z-suffix
  parse concern is likewise moot on 3.11+ (fromisoformat handles Z) — but the underlying
  silent-drop hardening (refactor: reorganize skills into 7 plugin packs for selective marketplace install #3) is kept regardless.
• agy [Critical] TypeError on null confidence → REFUTED: schema is confidence INTEGER NOT NULL CHECK(confidence BETWEEN 1 AND 10); Pydantic confidence: int = Field(ge=1, le=10). Never None.
  (Claude code-reviewer independently confirmed.)

Actionable NIT (must fix — convention cleans all NITs)

• Dead now param in score() — remove (implies time logic that isn't there).
• _slug_title double-prefixes (bash-bash-cd-git) — strip redundant prefix.
• SKILL.md LESSON_CONFIDENCE placeholder {{5-9}} → {{5-10}} (recurrence +1 ceilings at 10).
• "唯讀 lessons table" overstates — tighten to "只讀 lessons 資料列；schema init 由 init_db 寫入".
• Add high-value test assertions: union-find transitivity (AB,BC ⇒ one cluster), _jaccard
  partial-overlap ratio, boundary avg_confidence == 7.0 passes.

Voices unavailable

• None (all 3 active; Codex test-exec blocked by sandbox but review completed).

[howie]

Round-2 Mob Review — PR #173 (convergence)

Voice verdicts (round 2, on the fix commit)

• Claude (code-reviewer + silent-failure-hunter + pr-test-analyzer + comment-analyzer): LGTM
  — all round-1 findings verified resolved; pr-test-analyzer mutation-checked the new watermark
  resurface tests (an inverted > in score() now fails DT-017/DT-015). One new NIT (non-dict
  watermark JSON → AttributeError) — fixed in cc83f55.
• Codex: LGTM — no substantive findings (both rounds).
• agy / Gemini: NEEDS_CHANGES — all items refuted or documented (see below).

agy round-2 items → disposition

• [Critical] null confidence → int(None) crash → REFUTED (3rd time): schema
  confidence INTEGER NOT NULL CHECK(BETWEEN 1 AND 10) + Pydantic Field(ge=1, le=10). Never None.
• [Important] str(None) → "None" in clustering → REFUTED: insight/key are NOT NULL;
  nullable skill is already guarded by if m.get("skill").
• [Critical] scan-limit truncation + watermark skip → documented limitation: only triggers at

  2000 in-window lessons (current total ~233); surfaced via HarvestResult.truncated + CLI WARN;
  docstring on run_distill now records it. No SQL-paging redesign at current scale.

• [NIT] local import sys → addressed by moving it to load_watermark top.

agy reviews the diff without full-repo schema context, so it re-raises the same nullability
false-positives each round. Per this skill's "single-voice Critical → empirically refute", these
are dropped with evidence rather than cargo-culting a wrong fix.

Convergence

Functional convergence: Claude LGTM + Codex LGTM; agy's remaining objections are
evidence-refuted false-positives or a documented (warned) limitation. All real findings from both
rounds are resolved. 27 distill tests + 396 full mycelium suite + ruff/mypy/pylint/bandit green.