refactor(a11y-macos): move the name/description precedence into mapping

[xxx]

Fixes #262.

The macOS reader decided a node's name and description inside walk — cfg(target_os = "macos"), over a live AXUIElement with no constructible stand-in — so none of the rule could be tested off macOS. The issue's point was that the gate at its heart is behaviourally invisible: hoist the AXDescription read above the titled/untitled decision and every output stays byte-identical while every titled node with help text pays an extra cross-process round-trip. Only a read count catches that.

What moved

mapping::labels(title, read_description, read_help) -> (name, description) — the issue's form 2. The title is a value; the other two are FnOnce readers, so the extraction preserves the reads exactly (same attributes, same order, same counts) and a test can assert which attributes went unread.

mapping::node_name(title, read_description) holds the AXTitle-else-AXDescription precedence that walk, set_value and invoke each spelled out inline. labels delegates its name slot to it rather than restating the rule, so the walk and the two fingerprint sites cannot rank the two attributes differently and reject an element that never moved. The reader is one-shot: whichever slot spends the AXDescription read, the other finds it gone.

normalize_description moves with the decision, which brings the "description repeats the name" drop under test off macOS for the first time.

Whitespace

Pinned as it was, not trimmed. read_label folds "" but not " ", matching nonempty in the Windows and Linux readers. The Android reader deliberately counts a blank label as absent — so this is a live divergence between backends, not a rule this crate settles, and the test says that rather than claiming a convention.

Verification

Ablations, each run and restored:

Ablation	Tests that go red
hoist the AXDescription read above the decision	a_titled_node_with_help_never_reads_its_description only
retry the help fallback on a label that normalizes away	the two ..._without_retrying_the_description tests only
inline normalize_description lookalike: blankness, no trim	the combination table only
un-delegate labels and trim its own title arm	the_fingerprint_name_and_the_walked_name_agree_on_every_combination

Also: cargo test --workspace green; cargo clippy --workspace --all-targets --locked -- -D warnings clean; cargo clippy --target x86_64-apple-darwin -p glass-a11y-macos --all-targets --locked -- -D warnings clean, which type-checks the cfg(macos) reader from Linux; cargo fmt --all --check clean.

Not run here: crates/glass-macos/tests/a11y.rs, the on-box output check — it needs a TCC-granted Mac. Unchanged by this branch.

No CHANGELOG entry: internal refactor, which the file's own maintenance note excludes.

Known limits

• labels' two reader parameters are the same type, so transposing them at the call site compiles. One call site, and each closure names its attr:: constant on its own line, so this was judged not worth a newtype.
• "half the AxTarget fingerprint" is imprecise — name is one of three facts the gate checks — but the phrasing is inherited from glass-core/src/accessibility.rs and reader.rs, so it is left consistent rather than fixed in one place.

Follow-ups found while reviewing, not fixed here

• read_label's "" fold (reader.rs:370) and read_subrole's gate (reader.rs:411) are the same "delete it and only a read count changes" invariant this issue was filed about, and are still cfg(macos)-only.
• A whitespace-only name is reachable on macOS, Windows, Linux and iOS, and it displaces a real label: no name: selector reaches it, and outline::is_scaffolding stops eliding the node.
• walk reads attr::VALUE as a string (reader.rs:477) and gather_states reads the same attribute as an i64 (reader.rs:621), so every checkbox/radio/switch pays two round-trips on one attribute.