feat(hygiene): hygiene conductor — phase 2 modes (pre-scan + delegate) (#89)

[Jammy2211]

Closes #89 (phase 2 of the hygiene conductor).

What

Phase 2 gives the four non-perf hygiene modes real behaviour: each is a cheap,
read-only local pre-scan that delegates the full audit + any execution to
the owning skill. The conductor never runs a heavy audit and never mutates a repo
(consistent with its charter and the phase-1 boundaries). perf stays staged for
phase 3.

Modes (live)

Mode	Pre-scan (kind)	Delegates to
tidy	git debris — stale branches / stashes / [gone] refs / dirty checkouts (debris)	/repo_cleanup (Brain)
noise	none — needs a pytest + workspace-script run (advisory)	/cli_noise_clean (Heart)
deps	capped/pinned specifiers in library pyproject.toml (surface)	/dep_audit (Heart, hits PyPI)
docs	docs/api/*.rst + currentmodule counts, 3 doc repos (surface)	/audit_docs (Heart, imports)

noise/deps/docs route to read-only PyAutoHeart observation skills —
measurement lives in Heart; hygiene pre-scans, prioritises and routes.

Design note (a refinement of the approved boundary)

All four skills are pure agent-run markdown (three Heart-owned), so a stdlib/bash
conductor genuinely cannot execute them. The modes are therefore pre-scan +
route, not "run the audit". Pre-scans are typed so the ranking is honest:

• debris (tidy) — directly-removable items; a real, rankable count.
• surface (deps/docs) — only sizes the audit; the count is not a
  problem count (real findings emerge when the delegated skill runs).
• advisory (noise) — no cheap local signal.

Only debris drives the default ranking, so the no-arg run recommends tidy
when debris exists, else prompts the periodic audits. This strengthens the
"skills stay the executors" boundary rather than departing from it.

Changes

• agents/conductors/hygiene/hygiene.sh — the four pre-scans, the pre-scan-kind
  model, honest default ranking, extended --json HygieneDecision
  ({mode,kind,status,count,summary,delegate}).
• agents/conductors/hygiene/AGENTS.md — modes table + pre-scan-kind doctrine.
• skills/hygiene/hygiene.md — veneer updated (only perf staged).
• skills/repo_cleanup/SKILL.md — tidy is the front door; the skill stays the executor.
• tests/test_hygiene_conductor.py — new: hermetic contract tests (JSON
  structure, per-mode round-trip, exit codes, help) with PYAUTO_ROOT pointed at
  an empty temp dir.

Autonomous-ship gate (--auto, supervised)

• Tests — PyAutoBrain 34 pass (30 + 4 new). Downstream libs n/a.
• Smoke — n/a (Brain reasoning/org repo, no downstream script surface).
• Review — CLEAN, risk flags none.
• Heart — organism YELLOW, no launch ack; run parked at supervised ship
  sign-off and shipped on explicit live human go-ahead.

Validation checklist (reviewer)

• bin/pyauto-brain hygiene ranks tidy first when debris exists; deps/docs shown as surface.
• hygiene {tidy,deps,docs,noise} each emit real pre-scan + delegate; --json valid.
• hygiene bogus exits 2; hygiene --help prints the usage block.
• perf still reports staged (phase 3).

🤖 Generated with Claude Code

https://claude.ai/code/session_01Tw3EwV55k6VzxorYng3Kfn