Migrate the CLI from argh to cw

[thorwhalen]

Replaces argh (LGPL-3.0-or-later) with cw (MIT, zero runtime dependencies). Nothing a user sees changes.

The proof

misc/cli_golden_py310.json and misc/cli_golden_py312.json were recorded from the argh CLI before any source edit (their own commit, first in the branch) and replayed after each wave:

stage	3.12	3.10	--help
Wave 0 — from cw import compat as argh	18/18 identical	—	no changes
Wave 2 — cw.dispatch(run, argv)	18/18 identical	18/18 identical	no changes

--help diff: none — asserted, not eyeballed: strict_help=True compares the full help body, and python -m cw.testing diff-help prints no --help changes. No expect_diff entries. 38 passed (37 pre-existing + the new parity test).

The corpus — 18 cases, in misc/cli_cases.txt

• tier 3: --help. grub is one command, so this is its whole grammar;
• the no-query interactive prompt — grub's headline behaviour, recorded with stdin closed: banner, one prompt, exit 0 on EOF;
• searches through every flag, in both spellings argh derives — --method/-m, --chunk/-c, --extensions/-e, --snippets, and --n, the long form argh makes for a one-character parameter name;
• a multi-word query as separate argv items, which is what *query catches;
• grub's own bad-source path, which exits 1, not argparse's 2;
• five usage errors, all still exiting 2.

The argv handoff, which is the one thing that could have gone wrong

argh took it as a keyword (argh.dispatch_command(run, argv=argv)); cw takes it as the second positional (cw.dispatch(run, argv)). main still returns an int and still returns 0 on success, so test_cli_main_entry_point's main([...]) == 0 holds and the console script's exit code is unchanged.

One golden per CPython version

argparse is stdlib and rewrites its own text between versions: 3.12 stopped listing nargs='*' positionals among "the following arguments are required". That is the only difference between the two recordings, it predates this migration, and argh produces it identically — verified against plain argparse on 3.10/3.11/3.12/3.13 with neither argh nor cw in the picture. A single golden asserted across a matrix would fail for something nobody caused, which is the fastest way to teach a team to ignore a red parity test.

The 3.10 golden is recorded from argh too — a throwaway 3.10 venv over a worktree pinned at the pre-migration commit — so it is a migration proof on both matrix versions, not a baseline taken after the fact. An unrecorded version fails loudly with instructions rather than skipping.

It cannot rot

tests/test_cli_parity.py replays in the ordinary test run, so a future cw release cannot silently change this CLI. Verified it can fail: renaming snippets to snippet turned it red naming exactly that flag.

Worth recording that the first attempt at a negative test — forcing convention=cw.MODERN — did not fail, and that is correct rather than alarming: for a single command whose options are all keyword defaults, cw's two conventions produce identical grammar. A negative test that cannot distinguish anything is worth catching before it is trusted.

Two cw defects this migration found

Both fixed in i2mint/cw#34, both found here:

1. cw.testing inherited the recorder's stdin. grub's no-query REPL therefore recorded a timeout at a terminal and an EOF under CI — the same case, two different "facts". That made goldens non-reproducible between a developer's machine and CI, which is the exact failure mode RECORDING_ENV exists to prevent.
2. Windows' .exe was not scrubbed. argparse takes prog from basename(sys.argv[0]), so the runner reported usage: grub.EXE. (This one bit the sibling opsward migration first — 22 of 31 cases red.)

Also

• dependencies: argh → cw>=0.1,<0.2 (the fleet-wide spelling).
• rg -n 'argh' grub/ returns nothing.
• tests/fixtures/corpus/ — four one-line documents, well separated by topic, so a search case has a single unambiguous top hit.
• misc/README.md documents the harness: how to read a red result (you changed the CLI / scikit-learn moved a score / you changed the CLI by accident), how to re-record, and the four rules for adding a case.

One case was written and deliberately removed: --extensions .md over a .txt-only corpus raises ValueError: empty vocabulary out of scikit-learn, and the uncaught traceback names 17 absolute paths — unreplayable on another machine, and it names the dispatcher's own frames, so it could never have survived this migration regardless. Worth knowing that rough edge exists; it is not this PR's to fix.

Checks

Validation 3.10 · Validation 3.12 · Windows Tests — all pass.

Windows deserves a note: it was red before cw 0.1.1 (usage: grub.EXE ... on every case that prints a usage line) and is green after, with no expect_diff and nothing skipped. The whole corpus — exit codes, stdout, stderr, usage lines and the full --help body — is asserted on all three.