[thorwhalen]

Closes #5

Swaps the LGPL argh dependency for cw (MIT, zero runtime dependencies).

The change

# before
parser = argh.ArghParser(prog="lookbook")
argh.add_commands(parser, [curate, list_plugins, list_recipes, serve, mcp])
argh.dispatch(parser)

# after
COMMANDS = [curate, list_plugins, list_recipes, serve, mcp]
parser = cw.mk_parser(COMMANDS, prog="lookbook")
return cw.run(parser)

prog was already pinned, so it stays pinned — the python -m lookbook and
console-script surfaces remain byte-identical to each other, as they were.
cw.run returns the exit code where argh exited by itself, so main()
returns it (the console-script shim does sys.exit(main())) and the __main__
guard raises SystemExit(main()).

pyproject.toml: argh>=0.30 → cw>=0.1.1,<0.2. README's file-map line updated.

--help diff: empty

14 argv vectors were recorded from the argh implementation with
cw.testing.characterize before any edit — the top-level --help and -h,
all five subcommand --helps, the no-argument case, three non-zero-exit cases,
and the real output of list-plugins and list-recipes. Replayed against cw:

interpreter	result
CPython 3.12, cw from source	14/14 identical under --strict-help
CPython 3.10	grammar verified identical (see below)

--strict-help compares the full --help body, not just exit codes and the
usage line. Nothing moved: not the subcommand table, not the docstring-derived
descriptions, not the inferred short flags, not the nargs='*' rendering of the
Sequence[str] parameters, and not argh's quirk where --scorer and --selector
both lose their -s short flag to the collision between them.

The committed test replays non-strictly, because CPython itself rewrites
argparse's option column between versions and this repo's matrix spans several.
Separately verified that CPython 3.10 and 3.12 render this CLI's shapes
identically, and that published cw 0.1.4 (what CI will resolve) renders them
identically to the source tree — so the strict result above is what CI exercises.

What is now pinned

tests/test_cli_surface.py (7 tests) + tests/cli_goldens/lookbook.json:

1. The whole grammar, replayed against the argh recording.
2. Exit codes. Proven load-bearing: removing the return and the
   raise SystemExit makes 4 of these tests fail while all 159 pre-existing
   tests still pass. That is the failure mode invisible to every other test.
3. No arguments → usage on stdout, exit 0 — argh's behaviour, which plain
   argparse with a required subparser does not have.
4. COMMANDS matches what the help text advertises.
5. A hygiene test asserting the golden names no absolute path.

One case is deliberately not in the golden: list-recipes's real output
reads YAML profiles out of the user's app data folder, so a committed recording
would assert a property of the recording machine.

Tests

159 passed, 3 skipped before, 166 passed, 3 skipped after. ruff check clean.