fix(test): confine the RLIMIT_FSIZE atomic-write repro to a child process

[Mikola Lysenko (mikolalysenko)]

What

cargo test --workspace intermittently (on this machine: reliably) exits 101 with zero failed tests: utils/fs.rs's atomic_write_failed_stage_write_errors_and_keeps_target (added in #223) caps RLIMIT_FSIZE to 256 KiB and ignores SIGXFSZ process-wide during its window. #[serial] only serializes against other #[serial] tests — every concurrently-running sibling test, and the libtest harness itself (io error when listing tests: … FileTooLarge), that wrote >256 KiB during the window died with EFBIG and aborted the whole core-lib binary mid-run.

Why not just resize

The payload must stay within tokio's single 2 MiB write chunk: the bug this test pins is write_all buffering the whole payload and returning Ok before the background write hits EFBIG, with sync_all storing the error instead of returning it. A multi-chunk payload surfaces the error directly and stops exercising the swallowed-error path — so the cap can't be raised out of siblings' reach.

How

The parent arm re-execs the test binary filtered (--exact) to this one test with a marker env var selecting the capped body — the rlimit's blast radius is now a single-test child process. An anti-vacuity assert on the child's 1 passed summary keeps a future rename from turning the child run into a filter-matches-nothing no-op.

Validation

• RED-guard cycle: reintroducing the guarded bug (removing flush()) fails the confined test; restoring it passes.
• cargo test --workspace: 0/3 green before the fix → 2/2 green after on the machine that reproduced it; -p socket-patch-core --lib 2526/2526.
• Found while gating feat(get): --mode hosted|vendored with installed-version narrowing of advisory fan-outs #226 (which is unaffected — it touches no core file); root-cause narrative also in that PR's description.

🤖 Generated with Claude Code

---

Note

Low Risk
Test-only isolation of an existing Unix regression; production atomic-write code is unchanged.

Overview
Stops cargo test --workspace from aborting with exit 101 and zero failed tests when the Unix atomic-write EFBIG regression test is running.

atomic_write_failed_stage_write_errors_and_keeps_target still caps RLIMIT_FSIZE (and ignores SIGXFSZ) to pin the swallowed-sync_all path, but that cap is now applied only in a re-exec’d child filtered to this one test. #[serial] is dropped because it never protected concurrent non-serial siblings or the harness. The parent asserts child success and that stdout contains 1 passed so a rename cannot make the filter a no-op.

^{Reviewed by Cursor Bugbot for commit 183eaed. Configure here.}