chore: build the binaries plural from whole words, and document releasing

[max-sixty]

Two maintenance items found while cutting 0.4.1.

pre-commit run --all-files rewrote collect.rs's format!("{n} binar{}", …) to binary{}, reintroducing the "1 binaryy" output bug #82 had just fixed. binaries_phrase now selects the whole noun — format!("{n} {}", if n == 1 { "binary" } else { "binaries" }) — which is byte-identical output (the existing unit test pins 0, 1 and 2) and leaves typos nothing to flag. The first draft of this PR allowlisted binar in .typos.toml instead; fixing the producer is the better half of that trade, since the allowlist entry would have suppressed the token repo-wide to protect one call site that didn't need to emit it.

CLAUDE.md also gains a Releasing section. The process — bump the version, open a chore: release X.Y.Z PR, tag the squash commit, and let the tag push trigger release.yaml's Trusted-Publishing cargo publish — lived only in git history and that workflow's header comment, so cutting 0.4.1 meant reconstructing it from the previous release commits.

CLAUDE.md's pre-commit sentence is rewritten to describe the hook set rather than claim parity with CI. #97 adds pre-commit run --all-files to ci.yaml's lint job, so any claim about what CI does or doesn't gate would go stale on whichever of the two lands second.

cargo test (117 unit + 44 functional), cargo clippy --all-targets, and pre-commit run --all-files are all green, the last now leaving the tree clean.

Note for #97: its verification section records pre-commit passing on main, which was true when it was written on 2026-08-28 but stopped being true when #82 introduced the binar stem on 2026-09-16. Its new CI step would fail on today's main. This PR removes that blocker.

This was written by Claude Code on behalf of max-sixty

[cargo-affected-bot]

The typos entry is correct and it does silence the only hit — I reproduced both directions against this tree (typos 1.50.2): with the entry the repo is clean, and with it removed src/collect.rs is the sole error, binar should be binary. Every claim in the new Releasing section also checks out against release.yaml and history: the tag filter **[0-9]+.[0-9]+.[0-9]+* matches vX.Y.Z, the last five releases all went out as chore: release X.Y.Z PRs, and nothing in that workflow creates a GitHub release.

Two things worth deciding before this lands.

The stem is avoidable, which would make the allowlist entry unnecessary. binaries_phrase in src/collect.rs builds the noun as format!("{n} binar{}", if n == 1 { "y" } else { "ies" }). Moving the split to the whole word — format!("{n} {}", if n == 1 { "binary" } else { "binaries" }) — produces identical output, and the existing binaries_phrase_splits_the_noun_not_the_suffix test pins all three cases, so the change is covered. That leaves nothing for --write-changes to rewrite and no .typos.toml entry to carry. As written, the entry allows binar in every file in the repo, and the format string that produced #82's regression is still sitting there for the next --write-changes run to find — the hazard is contained by a config line rather than removed. Happy to push that as a commit on this branch if you want it: the format line and the doc comment above it that explains the split, plus dropping the .typos.toml hunk.

#97 is the other half of this sentence and disagrees with it. That PR adds a pre-commit run --all-files step to ci.yaml's lint job — which runs typos — and its body says explicitly "This PR fixes the gap rather than the sentence, so CLAUDE.md needs no edit." Both are open, so whichever merges second leaves the tree inconsistent: this PR's new sentence becomes false the moment #97 lands. Worth picking one before merging either.