fix(output): pluralize the count nouns in collect/run output

[cargo-affected-bot]

cargo affected collect printed exporting coverage maps for 1 binaryy... on any single-binary project — the plural was interpolated as a suffix (binary + y/ies), but the binary/binaries split lands mid-word, so the singular branch appended a stray y.

Tonight's rolling survey of tests/functional/diff_collect.rs turned up the same bug class in six more messages, so this PR now covers all of them rather than leaving the output half-fixed.

The wider defect

format_summary hard-coded tests, so the tool's single most-printed line read 1 tests to run whenever a selection came down to one test — which is the ordinary outcome of the workflow this tool exists for. It shows up verbatim in the repro transcript on #94:

cargo-affected: cache=hit-with-divergence selection=1/2 (50%) max_commits_ahead=1
1 tests to run (1 affected + 0 config + 0 new + 0 stranded, 1 skipped of 2 reachable-known):

Five more sites had it, including the found N tests across M binaries line whose binaries half this PR already fixed:

Site	Before at n=1
selection.rs format_summary	1 tests to run / 1 tests would run / 1 tests to recollect
collect.rs listing summary	found 1 tests across 1 binary
collect.rs update path	updating coverage for 1 tests (1 ranges)...
collect.rs store path	storing coverage for 1 tests (1 ranges)...
collect.rs done line	done. 1 tests, 1 ranges stored in ...
run.rs nextest handoff	running 1 tests with nextest

The pruned N test{s} and evicted N fingerprint{s} messages already pluralized correctly, which is what made the rest read as an oversight rather than a house style.

Changes

collect::plural_s(n) -> &'static str is the s in {n} test{s}, and every suffix-pluralized count in the codebase now routes through it — the six sites above plus the seven that already carried their own inline copy of the same conditional (pruned, evicted, the skipped test{s} warning and the completed test{s} bail in collect.rs, the missing-sha notice in selection.rs, the stale rows line in status.rs, and the leftover-staging-dir line in main.rs). binaries_phrase stays a separate function because its split is mid-word; its doc comment now points at plural_s for the contrast. The parenthesized commit(s) in collect.rs/run.rs/status.rs is a third convention and is deliberately left alone — it's already grammatical, and converting it would churn output this PR has no other reason to touch.

The regression test is selection::tests::summary_singular_test_count. Per tests/CLAUDE.md it was seen failing against the pre-fix expression:

thread 'selection::tests::summary_singular_test_count' panicked at src/selection.rs:620:9:
expected a singular noun for a one-test selection, got:
1 tests to run (1 affected + 0 config + 0 new + 0 stranded, 4 skipped of 5 reachable-known) — pass -v to list

The two pre-existing format_summary unit tests assert n=3 and n=4, so they pin the plural branch unchanged. Four functional assertions that pinned the ungrammatical form (tests/functional/diff_collect.rs ×3, tests/functional/run.rs ×1) are updated to the corrected wording.

Notes

cargo test --bins — 114 passed. cargo clippy --all-targets and cargo fmt --check clean. The functional suite needs cargo-nextest and llvm-tools, neither of which is installed in the agent sandbox (that's what #96 addresses), so those four updated assertions are verified by CI here rather than locally.

This conflicts with #75 at one line — tests/functional/diff_collect.rs:302, where that PR tightens the same assertion to include the category breakdown. Whichever lands second needs 1 tests to run (1 affected → 1 test to run (1 affected. Nothing else overlaps; #49, #81, #92 and #41 touch disjoint hunks.

Folded into this PR rather than opened as a new one, since it's the same concern in the same output surface and the queue is long enough already.

[cargo-affected-bot]

Two suffix-pluralization sites still carry their own inline if n == 1 { "" } else { "s" } and don't route through plural_s, so the sweep the description calls complete ("every suffix-pluralized count now routes through it") is two sites short:

• src/collect.rs, in the block guarded by if skipped > 0 — let s = if skipped == 1 { "" } else { "s" }; eprintln!("{skipped} test{s} produced no coverage");. This is the same shape as the pruned and evicted copies the PR did convert, and it sits ~20 lines above the updating coverage for … hunk.
• src/status.rs, in the reach.missing block — the stale rows: {stale_rows} (anchored at missing sha{}) line ends with if reach.missing.len() == 1 { "" } else { "s" } as its last format arg.

Both are already grammatically correct today, so this is consolidation rather than a bug — but leaving two copies of the predicate behind is exactly the drift that let binary + y ship in the first place. Happy to push a commit routing both through plural_s if you'd rather have the sweep land whole; both are outside this diff's hunks so there's no inline suggestion to apply.

Separately, note: {} commit(s) since collect (src/run.rs, src/status.rs, src/collect.rs) uses a third convention — the parenthesized (s) — for the same problem. Not worth folding into this PR, but it means "every count noun is consistent now" isn't quite true yet either.

The plural_s / binaries_phrase split itself is right, and the doc comments pointing at each other are what make the mid-word case hard to re-break. summary_singular_test_count pins the site that mattered, and the pre-existing n=3/n=4 assertions keep the plural branch honest. CI is green on f8ba8c1 across all three platforms, so the four updated functional assertions check out.

[cargo-affected-bot]

Pushed in b522b4e — and the sweep was three sites shorter than the review counted, not two. Beyond collect.rs's skipped test{s} warning and status.rs's stale rows line, the same inline predicate was still live in three more places:

• src/collect.rs — the refusing to overwrite stored coverage bail, ~15 lines below the skipped one, spelling the conditional inline as a format arg rather than binding it to let s.
• src/selection.rs missing_shas_notice — let plural = if missing.len() == 1 { … }, feeding the collect_sha{plural} in the divergence notice. This one is the upstream of the status.rs site the review found: status.rs prints the notice and then appends its own separately-computed sha{} suffix, so the two copies were computing the same predicate over the same reach.missing a line apart.
• src/main.rs clean — removed {swept} leftover staging dir{s}.

All five now call plural_s, so grep -rn 'else { "s" }' src/ is empty and the function is the only place the predicate is written. Rendered output is byte-identical at every site — plural_s is the same conditional — so no test expectations move; cargo test --bins is 114 passed, cargo clippy --all-targets and cargo fmt --check clean.

Agreed on leaving commit(s) alone. It's grammatical as written, and converting it would rewrite three more output lines for style in a PR whose reason to touch output is a bug. The PR description now says which convention lives where instead of claiming the codebase is uniform.

Why the three extra sites didn't match the review's grep shape

The review found the two that bind the predicate to a let s on its own line. The other three vary the surface: selection.rs binds to let plural, collect.rs's bail passes the conditional straight in as a positional arg to bail!, and main.rs sits in a subcommand body rather than the collect/status output paths the sweep walked. Same predicate, three spellings — which is the drift argument the review made, one layer down.