test(eval): avoid writing executable runtimes during parallel tests

[aheritier]

🤖 Automated implementer agent — this comment was posted by the implementer bot from Docker Agentic Platform, not by a human developer

Fixes #4520

Change

Write one immutable shell runtime fixture before the parent t.Parallel() barrier. Each parallel case symlinks that executable and writes only non-executable response/exit-marker files. This is test-only: the existing 12-case table, every assertion and both levels of parallelism remain unchanged. No production changes, retries, sleeps or relaxed assertions.

Diagnosis

The original parallel cases write executable scripts immediately before executing them. Natural reproduction surfaces Linux ETXTBSY (text file busy). An inspect execution error takes the intentional production fallback, leaving inspect-args absent and sometimes the Dockerfile's COPY unpinned; build execution errors expose the errno directly. The recovered CI log includes both no_manifest_digest and default_image failures.

A concurrent fork temporarily inheriting a sibling executable's writer descriptor is a plausible explanation, not syscall-proven. Prior diagnostic logging exposed inspect ETXTBSY; strace changed scheduling and did not reproduce it. The fix removes concurrent writes to the executed fixture inode.

Validation

Prior investigation (separate runs, not fresh PR-arc measurements):

Fixture/revision	Failing parent iterations	Failing subtests
Original, failing 6bf72739	30/100	37/1,200
Original, main 28569802	34/100	39/1,200
Candidate, both revisions combined	0/1,200	0/14,400

Fresh PR-arc validation on main 285698023342389523089b3a36171eb88fb8dc40:

Command	Result
Original target: go test -json -race -shuffle=1791198938208945572 -count=100 -run '^TestBuildEvalImagePinsLocalAgentDigest$' ./pkg/evaluation	33/100 parent failures; 41/1,200 subtest failures; explicit text-file-busy errors
Candidate: same focused command	0/100 parent failures; 0/1,200 subtest failures
Candidate: focused -race -shuffle=on -count=100	0/100 parent failures; 0/1,200 subtest failures; seed 1791278474244450004
Candidate: whole evaluation package -race -shuffle=1791198938208945572 -count=3	Passed
task build, task test, task lint	Passed; lint 0 issues/no custom-cop offenses; module tidiness clean
CGO_ENABLED=1 task test-race	Passed, full repository

Focused commands source .env.test and set CGO_ENABLED=1. Local Go1.27.0/Linux Ubuntu26.04 differs from CI Ubuntu24.04.5; Windows remains skipped, macOS untested. PR CI runs race tests only on main pushes, so these are local race results, not PR-CI results. Finite passes are not a suite-wide flake-freedom claim.

Other observed flakes (not changed)

• Prior full task test first failed the unrelated pkg/cache/TestFileCache_dedupSkipsRedundantWrite modtime assertion, then passed on retry. Fresh unpatched cache control failed 17/20 repeats; this PR's fresh full task test passed first time. Cache source is unchanged.
• Initial fresh random whole-package race/count3 run (seed 1791278540484914705) failed once in unchanged TestEvalContainerForwardsOrderedFlavors/#00, flavors_test.go:37, executing its own per-case script: fork/exec .../runtime: text file busy. That fixture has the same parallel write-then-exec pattern. Target digest test passed all three repetitions. Clean unpatched-main controls did not reproduce the sibling failure (100 focused race runs, 10 same-seed package race runs); candidate same-seed package/count3 rerun and subsequent full race suite passed. The sibling's frequency and pre-existing attribution are not established by these samples. Left outside this one-test fix.

[Copilot]

Copilot review overview

🟢 Approval recommended

The focused test-only change preserves existing coverage while addressing the demonstrated race condition.

Review effort: Balanced
Findings: None

What changed in this PR

Stabilizes the parallel evaluation image tests by sharing a prewritten runtime fixture.

Changes:

• Creates the executable fixture before parallel execution.
• Uses per-test symlinks and non-executable response files.

File	Description
pkg/​evaluation/​build_test.go	Prevents concurrent executable writes that caused ETXTBSY failures.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.