tests(drift): cover the missing collect_sha, and assert the sibling case positively

[cargo-affected-bot]

tests/functional/drift.rs is the pair of scenarios that decide whether a collect_sha which is no longer on HEAD's lineage still anchors the diff. Both halves were asserted only negatively, so neither pinned the behavior its name claims, and both failure modes they exist to catch are ones nothing downstream reports.

The sibling half asserted nothing about selection. sibling_collect_sha_status_uses_selection checked that status printed neither would run all tests nor not in the repo. The empty-selection path prints neither either — it prints no tests cover the changed lines and no new tests and returns — so the scenario stayed green through a sibling sha that anchored nothing at all. That is silent under-selection, which tests/CLAUDE.md names as the one failure mode this tool cannot detect downstream. It now asserts the two halves positively: the cache stayed usable and stayed classified as diverged (cache=hit-with-divergence, which plan.rs reaches for any non-Equal relation, so a sibling misclassified as Equal fails the scenario rather than passing a looser cache=hit substring), and it produced a selection (tests would run). Reading combined_output rather than stdout is what reaches the summary line, which status writes to stderr.

The missing half had no scenario at all. missing_collect_sha_status_widens_to_all_tests covers the branch where the only collect_sha is gone: CacheMiss::NoReachableSha, which widens status to the full suite. Nothing else reaches it. The missing-sha notice on its own is already pinned positively by run_unions_affected_and_stranded_when_sha_is_missing, but only on the partial-divergence path, where a second sha survives and the lost one's tests come back as stranded; would run all tests is asserted positively nowhere else in the suite. Resetting alone only orphans the commit — the reflog still holds it and relation_to_head reports Reachable, which is the sibling case — so the scenario expires the reflog, prunes, and asserts the object is gone before running status, or it would silently exercise the other half of the pair. That prune sequence is the same one diff_collect.rs uses, comment included; folding both into a shared main.rs helper is worth doing but touches a second scenario file, so it is left out of a tests-only coverage change.

Verified against the live behavior rather than by reading: in the sibling scenario status reports cache=hit-with-divergence selection=3/3 (100%) and 3 tests would run (3 affected + 0 config + 0 new + 0 stranded, 0 skipped of 3 reachable-known). Flipping each new assertion to a string the output does not contain fails the scenario, so none is vacuous. cargo test is 45 passed / 0 failed; cargo fmt --check and cargo clippy --all-targets are clean.

Why the negatives alone could not fail

Reproducing the empty-selection path on a two-module project — fresh collect at HEAD, clean tree — status prints:

cargo-affected: cache=hit-exact selection=0/3 (0%)
coverage database: target/affected/coverage.db
...
no changes since the newest collect_sha and no new tests — nothing would run

Neither would run all tests nor not in the repo appears, so both pre-existing assertions pass over a selection of zero. That is what the selection assertion closes, and it is why it is needed alongside the cache one rather than instead of it: cache=hit-with-divergence catches a widening-by-cache-miss and a sibling misclassified as Equal, tests would run catches an empty selection, and those are different regressions.

main has moved since this branch opened, so it carries a merge of main as well.

[cargo-affected-bot]

The test itself holds up. Traced the path it pins: with the only collect_sha pruned, check_shas_reachable puts it in missing with reachable empty, so plan yields CacheState::Miss(CacheMiss::NoReachableSha) — whose reachability() still returns Some, which is what makes status print the notice on stdout before the "no reachable collect_sha … would run all tests" line. Both asserted strings land on stdout, and the two together pin the branch: not in the repo can only come from a non-empty missing, so the generic "would run all tests" can't be satisfied by the no-DB or fingerprint-miss messages instead. The cat-file -e <sha>^{commit} guard mirrors relation_to_head exactly, so it fails for the same reason the classifier would.

One adjacent thing, outside this diff — the doc comment on the function this test now pins is stale. missing_shas_notice documents its parameter as "will rerun as 'new'" for run/collect --diff, "would rerun as 'new'" for status, but no caller passes 'new': run.rs passes "will rerun as 'stranded'", status.rs passes "would rerun as 'stranded'", and collect.rs passes a third shape the doc doesn't mention at all ("will be rerun and re-anchored at the new HEAD"). Not for this PR — it's tests-only and should stay that way — but it's a candidate for the rolling prose sweep in #88.

[cargo-affected-bot]

The added scenario earns its place: CacheMiss::NoReachableSha — every collect_sha gone, so status widens to the full suite — is reached by no other test, and "would run all tests" is asserted positively nowhere else in the suite. Two findings on top of that, and one thing to note.

The missing-sha notice is already pinned positively. run_unions_affected_and_stranded_when_sha_is_missing asserts combined.contains("not in the repo") under the comment "The missing sha is announced, not silently dropped", and reaches that state through the same reflog-expire → gc --prune=now → cat-file -e sequence. So "nothing pinned the wording — or the emission — of the path those negatives are defined against" isn't accurate, and the PR body repeats it. What is genuinely uncovered is narrower: that scenario leaves a second sha reachable, so it exercises the partial-divergence/stranded path, whereas this one is the only test of the branch where every sha is gone. Stated as-is, a later reader who finds the diff_collect.rs assertion concludes this scenario is redundant and deletes it — losing the one branch it actually covers.

cache=hit also matches cache=hit-exact. plan.rs classifies any non-Equal relation as HitWithDivergence, so the sibling sha here is hit-with-divergence. A regression that reclassified a sibling as Equal would keep the substring and leave the scenario green — which is inside the class of silent misclassification the new assertion was added to close.

One to note rather than change here: the fifteen-line "make this commit genuinely unreachable" block is now duplicated across two scenario files, reasoning comment included. A main.rs helper is the natural home for it, but that touches diff_collect.rs and is beyond a tests-only coverage PR.