modern_decode: read abstract sequence interfaces as several values

[thorwhalen]

The problem

argh recognises list and nothing else, so a parameter annotated names: Sequence[str] became one positional taking one token:

$ tool check scale eq lut3d
usage: tool [-h] {check,...} ...
tool: error: unrecognized arguments: eq lut3d

The failure is quiet in the way that costs time: the parser builds, --help looks right, the command runs, and the error points at the value rather than at the annotation. And it penalises exactly the annotation style that prefers collections.abc interfaces over concrete containers — Sequence[str] means the same thing to a reader as list[str], and meant something else to the parser.

Measured before the change:

annotation	variadic?
list[str], List[str], list	✅
Sequence[str], Iterable[str], tuple[str, ...], Sequence	❌

The change

modern_decode now covers Sequence / MutableSequence / Iterable / Collection / Set / MutableSet, plus tuple / set / frozenset, subscripted or bare.

argh_decode is untouched. It is argh's if-chain if-branch for if-branch, and widening it would break the compatibility contract that is its whole purpose — modern_decode is the seam for exactly this, and its fall-through composition made the change three lines. A test asserts argh_decode still returns {} for every new spelling and still returns what it always did for list[int].

Three decisions

• str and bytes must never be read as sequences, and this avoids it by construction rather than by a special case: matching is on typing.get_origin, and get_origin(str) is None. Both are registered Sequences, so an issubclass test would have read every string parameter as variadic and broken every existing CLI. Tested directly, because it is the trap this change lives inside.
• A homogeneous fixed-length tuple gets a fixed nargs — tuple[int, int] → {'nargs': 2, 'type': int}. It is the one case where the annotation carries a count, and it makes --size W H fall out of the signature.
• A heterogeneous tuple gets no inference at all. add_argument has a single type and there is no honest one to choose, so tuple[int, str] → {} rather than guessing one and converting the other wrongly.

Tests

952 passing (was 932). 20 new: the abstract interfaces subscripted and bare, the str/bytes trap, both tuple cases, Optional[Sequence[int]], and the argh_decode-is-unchanged guard.

Second commit is pure formatting of two test files ruff format had drifted from — split out so the functional commit contains only the functional change.

Where it came from

Adopting cw as the CLI for looks, whose house style annotates with collections.abc interfaces throughout. cw was the right package — MIT, dependencies = [], and import cw verified to pull in nothing beyond stdlib — it just could not read the annotations the calling code was already written with.

https://claude.ai/code/session_01CnXLihXKG9jfvygHvgKVpX