GeezerVison

Summary

Fixes the coverage-job flake first seen on PR #139 (failing run): update::download::tests::sanity_exec_rejects_wrong_program_and_honors_strictness died with

VerifyFailed("downloaded binary failed to execute (wrong architecture?): Text file busy (os error 26)")

Root cause — the classic Unix fork/exec fd-inheritance race, not a broken test assertion: between a sibling thread's fork() and its exec(), the child briefly inherits every open fd, including the write fd on a script/binary written moments earlier. Exec'ing that file during the window fails with ETXTBSY. The parallel test binary forks constantly, and llvm-cov instrumentation widens the window — which is why it bit coverage while the identical test passed in test/test-release of the same run. The race is equally real for production self-update in any multi-threaded process.

Changes (two commits, RED then GREEN)

1. test(update) — RED: adds sanity_exec_retries_when_binary_briefly_text_busy (linux-gated; other platforms don't reliably enforce ETXTBSY), which reproduces the race deterministically by holding a write fd on the staged script for 150 ms from another thread while sanity_exec runs. Also tightens the existing strictness test to assert each rejection's reason instead of bare is_err() — an ETXTBSY spawn failure previously masqueraded as the expected rejection on the assertion one line above where the flake surfaced. This commit is pushed alone so CI demonstrates the failure in the exact environment that flaked; see the first (red) run on this branch.
2. fix(update) — GREEN: sanity_exec retries the spawn on ErrorKind::ExecutableFileBusy (up to 10 attempts, 25 ms linear backoff, ≤ ~1.4 s worst case) instead of failing a fully SHA-verified download. Same dance Go's os/exec and cargo do. All other spawn errors still fail immediately with the existing message; the 10 s hang timeout still applies per attempt.

Verification

• macOS: full update::download suite passes with the fix (regression test correctly linux-skipped).
• Linux (CI): the branch's first push (test-only) is expected RED on the new test in test/test-release/coverage; the fix commit turns it GREEN — commit history preserves the evidence.

🤖 Generated with Claude Code