Migrate the CLI from argh to cw

[thorwhalen]

Closes #39.

argh is LGPL-3.0-or-later. cw is MIT, has no runtime dependencies, and reproduces argh's grammar by default — so this changes nothing a user sees.

The CLI surface: no diff

Recorded from the argh console script before any edit, replayed against cw after, with cw.testing:

19 characterized cases -> Counter({'identical': 19})
assert_replay(golden, strict_help=True)  ->  clean

strict_help=True asserts the full --help body, not just the usage: line — so a moved help column or a reflowed description would have failed. Nothing moved.

The 19 cases: no-args, top-level --help, --help for all eleven subcommands, and six error/edge vectors (estimate with no args, inventory --nosuchflag, scaffold, publish, plan, nosuchcommand).

python -m coact --help was checked separately and still prints usage: __main__.py — see "prog" below.

What changed

The three @argh.arg declarations became a config dict. This was the modern form, not the compat shim:

CONFIG = {
    "inventory": {"project": dict(nargs="?", default=".", help="Project root (default: current dir)")},
    "scaffold":  {"agents":  dict(nargs="+", help="Agent .md files / skill sources to wire")},
    "publish":   {"source":  dict(nargs="+", help="Tool refs ('module:function'), a skill dir/SKILL.md, or both")},
}

def main() -> None:
    raise SystemExit(cw.dispatch(COMMANDS, config=CONFIG))

The eleven verbs are now plain annotated functions with nothing CLI-specific attached — the CLI's particulars live in one table next to the command list instead of being scattered across three decorators.

estimate(*agents: str) needed no entry: VAR_POSITIONAL infers nargs='*' on its own, and the replay confirms the help is byte-identical.

prog is deliberately not passed. Hardcoding prog='coact' would have been an improvement to python -m coact --help (which currently reports usage: __main__.py) — and therefore a change to the surface this PR promised not to touch. Left to argparse.

raise SystemExit(...) is load-bearing. argh.dispatch_commands called sys.exit itself; cw.dispatch returns the code. Dropping the re-raise would silently turn every usage error into exit 0. There is now a test that runs main() and asserts SystemExit(2), so it cannot be dropped by accident.

The test that had to be rewritten

test_main_wires_every_verb monkeypatched cli.argh.dispatch_commands. That monkeypatch existed because argh genuinely could not be observed any other way — its output file is bound at import and its return value is discarded, so the only way to see what got registered was to replace the dispatcher.

cw builds a plain argparse.ArgumentParser, so the wiring is inspectable directly. The monkeypatch is gone, replaced by five real tests:

• the parser's subcommand set is exactly the eleven documented verbs;
• the four positional nargs shapes (? with default='.', +, +, *) asserted on the real actions, parametrized;
• two end-to-end dispatches through cw.dispatch(..., out=buf) — including inventory with no positional, which is what nargs='?' is for;
• the exit-code test above.

Net: the CLI is better covered after the migration than before, and none of the new coverage depends on the dispatcher's identity.

Also

• pyproject.toml: argh -> cw.
• The ai-assistant-command-mcp skill carried an explicit argh-migration stage 2 marker — "a house CLI adapter is being built ... this section will name it and show its idiom once it ships." It has shipped, so the marker is retired and Emitter C now shows the cw dispatch-table idiom.
• Stale argh mentions updated in COACT_SPEC.md, REUSE.md, and the coact-dev skill. CHANGELOG.md left alone — it is history.

Tests

417 passed, 4 skipped.