Battam1111/Myco

v0.7.2 — 2026-04-30 — 永恒删减 ratchet dims (mechanize the eternal-pruning discipline)

Zero R1-R7 surface deltas; 4 NEW lint dimensions (46 → 50, additive); zero subsystem changes; schema v2 unchanged (additive canon fields). This molt converts the v0.7.1-named public-API-deletion discipline from doctrine prose to mechanically-enforceable lint. Future drift is continuous-detected by the substrate's own immune system instead of episodically discovered via manual neat-freak audits.

Why this release

L0 P3 (永恒进化 — eternal evolution) is two-sided: growth + shedding. Through v0.6.x the substrate was strong on growth (additive verbs, new lint dims, expanded subsystems) but weak on shedding. The v0.7.0 incident exposed accumulated bloat that hid 4 fail-silent dims for 16 minor versions; the v0.7.0 reactive deletion broke the owner's MCP host within 2 hours (v0.7.1 hotfix). Without ratchets, every accumulation cycle requires a manual audit pass to surface (v0.6.16-style).

5-critic L0-P1-P5 fanout (second substantive use)

Per the v0.6.15 doctrine, 5 parallel Opus critics (one per L0 principle, narrow visibility scope each). 22 findings (1 P0 BLOCK + 13 P1 + 4 P2) reshaped the proposal materially:

• Round 1 SH3 (shipped) → Round 2 MB8 (metabolic) per saprotroph T1: runtime telemetry on substrate state ≠ publication-surface invariant.
• Round 1 MB8 (metabolic) → Round 2 PA6 (mechanical) per saprotroph T2: filesystem invariant ≠ agent-throughput pressure.
• Round 1 PA6 (mechanical) → Round 2 MF5 (mechanical/MF) per saprotroph T3: manifest/file-shape ≠ package-architecture purity.
• Round 1 SE5 → kept (saprotroph T4 misread; SE2 is orphan-detection, not canon-cited-paths — doctrine drift in homeostasis.md confirmed and corrected).
• SH3 import-time write → first-MCP-request hook per mycorrhiza T1 P0 BLOCK: read-only substrate compatibility (v0.7.0 incident lesson).
• L2 boundary.md amendment WITHDRAWN per saprotroph T5: would require R7 contract-bump path. Discipline kept as L3 dim docstrings + 1-paragraph homeostasis.md cross-reference.
• Recursion-cutter HARDENED per mycoparasite T1/T2/T3/T6: 4 new path patterns + 4 new canon keys + NEW multi-cluster compound trigger.

What changed

Group A — 4 new lint dimensions (~640 LoC source + ~320 LoC tests)

• MB8 (metabolic, MEDIUM) — shim-hit counter. Reads append-only .myco_state/shim_hits.json (written by shim's __main__.py CLI hook). Reports per-module hit counts + last-hit ages + sunset-gate thresholds (governance.shim_sunset_min_zero_cycles + _days). Enables the v0.7.1 telemetry-verified safe-deletion gate.
• PA6 (mechanical, MEDIUM) — repo-bloat detector. Walks tree using core.skip_dirs.should_skip_path (canonical exclusion list) + metrics.repo_size_excluded (substrate-level globs). Compares to metrics.repo_size_max_bytes (default 50 MB). MEDIUM at ≥80%, HIGH at ≥100%. Resolution path is agent-autonomous via myco excrete or fruit→winnow→molt; never owner-merge-gated (chytrid T3 / L0 P1 preservation).
• MF5 (mechanical, LOW) — generated-mirror integrity. SHA-256 bucket O(n) hash detection of byte-identical pairs across .claude/{agents,commands}/ ↔ <repo>/{agents,commands}/. Reports v0.6.11-documented mirrors as PENDING_BUILD_ARTIFACT_CONVERSION (LOW; v0.7.3 IOU). UNINTENDED_DRIFT pairs at MEDIUM.
• SE5 (semantic, LOW) — version-anchor freshness. Greps live agent-facing docs (docs/architecture/**, MYCO.md, README*.md, _canon.yaml, pyproject.toml) for v0.X.Y anchors > 3 minor versions stale. Excludes _archive/, _landed/, contract_changelog/. False-positive heuristic: historical-context tokens (shipped at / landed in / as of / since / pre- / post-) suppress.

Group B — risk_classifier recursion-cutter hardening

• _RECURSION_CUTTER_PATH_PATTERNS extended (mycoparasite T1, T2):
  • .myco_state/shim_hits.json + .myco_state/*.json (telemetry files closed against zeroing attack)
  • src/myco/mcp/** (the shim itself; deletion now HIGH-tier)
  • src/myco/boundary/mcp/** (the canonical launcher)
• _RECURSION_CUTTER_CANON_KEYS extended (mycoparasite T3): repo_size_max_bytes + repo_size_excluded + shim_sunset_min_zero_cycles + shim_sunset_min_zero_days (closes the threshold-bumping bypass).
• NEW _RECURSION_CUTTER_COMPOUND_CLUSTERS (mycoparasite T6): a craft whose path_allowlist simultaneously touches ≥ 2 of {state, shim, canon, classifier} clusters forces HIGH regardless of individual path classification. Defeats the compound 4-step attack: [_canon.yaml + .myco_state/shim_hits.json + src/myco/mcp/__init__.py] in one craft.

Group C — canon schema additions (additive within v2)

• metrics.repo_size_max_bytes: 52428800 (50 MB default)
• metrics.repo_size_excluded: ["notes/**", "docs/primordia/_landed/**", "docs/_archive/**", "docs/contract_changelog/_archive/**"] (rhizomorph T1 fix — protects substrate's own ingestion from being its own bloat-cause)
• governance.shim_sunset_min_zero_cycles: 7
• governance.shim_sunset_min_zero_days: 7

Group D — shim __main__.py telemetry hook (mycorrhiza T1)

Counter-write happens at the CLI entry (python -m myco.mcp), NOT at module import. Try-except wrapped: read-only substrate / missing canon / locked file → silent no-op. The substrate MUST always boot the MCP server (v0.7.0 incident lesson preserved).

Group E — L2 doctrine update

• homeostasis.md SE2 description corrected from "canon-cited numbers and paths" (stale v0.5.8 doctrine; current SE2 is orphan-integrated detection) to actual implementation. The "canon-cited paths drift" intent is now covered by SE5.
• New § "永恒删减 (eternal pruning) — the v0.7.2 ratchet quartet" cross-references the 4 dims under their cleaned-up categories.

Group F — gitignore hygiene

.myco_state/sessions.db / boot_brief.md / autoseeded.txt / decay_baseline.yaml / search_misses.yaml / excretion_counter.json / graph.json per-file list collapsed to .myco_state/** wildcard. Future state files (like shim_hits.json from MB8 telemetry) are excluded by default.

Break from v0.7.1

No backward-compat break. Schema v2 is unchanged (new fields are additive; old canons read fine — dim default to threshold defaults). 50-dim roster supersedes 46 (4 net new); existing dim IDs untouched. The new dims emit informational findings (exit_policy: metabolic:never, semantic:never keeps non-critical findings from gating). Shim ABI unchanged (telemetry hook is best-effort + invisible to consumers).

What did NOT change

• All 7 R-rules: identical text.
• All 7 subsystems: identical doctrine.
• 46 existing lint dimensions: identical roster, severities, fixability (the 4 new dims add to the roster).
• 20 verbs: identical manifest, CLI, MCP shape.
• system.llm_policy: forbidden: unchanged.
• system.governance.auto_evolve_* settings: unchanged.

Pre-flight gate verification

• ruff check src tests scripts — clean
• ruff format --check src tests scripts — 0 reformats
• mypy src/myco — 152 source files, no issues
• pytest -q — 1554 passed, 1 skipped (+22 from v0.7.1 baseline)
• myco immune — exit 0; new findings: 1 MB8 (telemetry working) + 11 MF5 PENDING_BUILD_ARTIFACT_CONVERSION (v0.6.11 IOU surfaced) + 6 SE5 LOW (next-sweep candidates)
• python -m myco.mcp --help — boots
• python -m myco.boundary.mcp --help — boots

Governing craft: docs/primordia/v0_7_2_eternal_pruning_ratchets_craft_2026-04-30.md.
Predecessor: v0.7.1 (myco.mcp shim revival hotfix), shipped 2026-04-29.
Successor (planned): v0.7.3 (<repo>/{agents,commands}/ build-artifact conversion via scripts/build_plugin.py extension; resolves the 11 MF5 PENDING findings).