Myco v0.5.4 — Dogfood-session patch (seven bugs fixed)

TL;DR: Yanjun asked the Agent to run Myco on the Myco repo as an end-to-end smoke test. Seven bugs surfaced, all seven are fixed, each pinned by a regression test. Two were critical — they prevented substrate-local plugins from working at all. No contract-surface change from v0.5.3.

Critical fixes

• Bug #6 — ramify --dimension was broken. The subparsers' dest="verb" collided with ramify's --verb <name> argument: the flag's default None overwrote the subcommand selector, so running myco ramify --dimension X --category mechanical --severity medium crashed with UsageError: unknown command: None. Every dimension / adapter scaffold invocation was broken. Renamed subparsers dest to _subcmd (private, non-colliding).
• Bug #7 — Substrate-local plugins never registered. The ramify-generated .myco/plugins/dimensions/__init__.py template contained {{__name__}} (double-braces intended for a .format call that never happened). As Python these resolved to {__name__} inside an f-string, making importlib.import_module try to import a module literally named {__name__}.local1 — ModuleNotFoundError, silently swallowed by the outer plugins/__init__.py's best-effort try/except. Result: every substrate-local dimension registered via ramify was invisible to default_registry(), graft --list, and immune. Fixed template. Added issubclass(obj, Dimension) guard so only real Dimension classes register. Added an end-to-end regression test that loads a fresh Substrate and asserts the dimension actually registers.

Smaller fixes

• Bug #1 — myco --version / -V now prints the version and exits 0 (was: "VERB required").
• Bug #3 — list flags accept natural multi-value: --tags a b c works (was: only --tags a --tags b --tags c). nargs="*" + action="extend".
• Bug #2 — hunger payload gains local_plugins.count_by_kind: {dimension, adapter, schema_upgrader, overlay_verb}. v0.5.3 CHANGELOG promised this; v0.5.3 implementation shipped only a flat count.
• Bug #5 — --json output carries a top-level findings: [...] array so Agents learn why an immune run exited nonzero (which dimension / category / severity / path) without a follow-up verb call.
• Observation #4 — new winnow gate G6_template_boilerplate. Before: a freshly-fruit-ed craft skeleton passed every shape gate by construction — defeating winnow's purpose as a "did the Agent actually do the craft work" signal. Now: bodies >40% fruit-template fingerprints trip G6. All three existing v0.5.x craft records (v0.5.0 / v0.5.2 / v0.5.3) still pass with zero violations.

Test coverage

499 passed (was 486 at v0.5.3). +12 new regression tests, +1 LOCAL1 dimension surfaced via substrate-local registration now visible in graft --list.

Upgrade

cd ~/myco && git pull && myco immune

No migration action required. Every v0.5.3 invocation still works. The local_plugins.count field remains for backward-compat; count_by_kind is additive. The G6 gate is additive to winnow.

Records

• Full CHANGELOG entry: CHANGELOG.md
• Contract changelog: docs/contract_changelog.md v0.5.4 section
• Dogfood session was the v0.5.3 post-release audit (no separate craft — the bugs were mechanical, not architectural)