ci(mutants): add --in-place to dodge cargo-mutants v27.0.0 #611 tmp-tree bug

[mcarvin8]

Summary

Both full mutation sweeps run after #37/#38/#39/#40 landed -- run 25807382458 and run 25812733214 -- died with the exact same upstream error after passing the actual mutation score check (missed.txt was empty in both):

ERROR cargo_mutants::lab: Worker thread failed:
  "/tmp/cargo-mutants-config-disassembler-<rand>.tmp/src/xml/cli.rs"
  is not a file

The worker thread crashes before cargo-mutants can write its summary, the step exit code is 1, and the workflow goes red even though the underlying score is 100%.

Root cause (upstream)

This is cargo-mutants v27.0.0 issue #611, a regression introduced by #557 in v26.0.0 (Dec 2025). cargo-mutants copies the source tree into a per-mutant scratch directory under /tmp using reflink::reflink, which preserves the source file's mtime exactly. On macOS, /usr/libexec/dirhelper periodically scans /tmp and unlinks regular files whose mtime is older than CLEAN_FILES_OLDER_THAN_DAYS (default 3); on Linux, systemd-tmpfiles does the same on a configurable cadence. Any file in the repo with sufficiently-old mtime ends up in the scratch tree as a reaper target, gets silently unlinked mid-run, and the next mutant's BuildDir::overwrite_file trips this check at src/build_dir.rs:96:

ensure!(full_path.is_file(), "{full_path:?} is not a file");

The bug is load-dependent: short runs that finish between reaper invocations don't trigger it. Our debug.log on both failing runs shows the failure immediately after the revert of parse_reassemble_args -> (None, None, true), which is consistently late enough in the run for the reaper to have already fired and unlinked src/xml/cli.rs (the file in our repo with the right combination of stale source mtime + position in cargo-mutants' deterministic mutant order).

The fix landed upstream in #613, merged May 11 2026, which bumps dest mtime to now() after every reflink and gives a clearer error message. But the latest tagged cargo-mutants release (v27.0.0, 2026-03-07) does not contain the fix, and taiki-e/install-action installs from tagged binary releases via cargo-binstall.

Workaround

The upstream issue explicitly recommends --in-place for users not ready to switch off v27.0.0: it bypasses the scratch-tree copy entirely and runs mutations against the workspace source files in the runner's checkout directory. On ephemeral CI runners that's exactly what we want -- the runner is thrown away after the job, so "mutating the actual source tree" has no downside. --in-place takes the same --in-diff / --file filters and produces the same report shape as the default mode.

Both the full and incremental jobs are updated:

• full: cargo mutants --no-shuffle --in-place
• incremental fast path: cargo mutants --no-shuffle --in-place --in-diff mutation.diff
• incremental test-only-diff fallback: cargo mutants --no-shuffle --in-place --file ...

Each command grows a single comment block linking back to upstream #611 / #613 so the next maintainer can drop the flag once cargo-mutants ships a fixed release.

Test plan

• Merge.
• Manually trigger Full mutation testing (gh workflow run mutation.yml -f full=true).
• Expected: no is not a file worker error, exit code 0, the workflow step turns green, and the mutants.out artifact still reports missed=0 with the same caught / timeout / unviable buckets as the prior post-refactor(xml/cli): use iterator-based loop in parse_disassemble_args #40 run.

[codecov-commenter]

Codecov Report

✅ All modified and coverable lines are covered by tests.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.