Evals binary_tests grader always reports 0/N on macOS: it shells out to timeout (GNU coreutils) and interpolates the test command as one argv token #2188

Observed on 7.54.0, macOS.

What

// skills/Evals/Graders/CodeBased/BinaryTests.ts:33
const result = await $`cd ${workingDir} && timeout ${Math.ceil(timeout/1000)} ${command} ${testFile}`
  .quiet().nothrow();

Two defects in one line:

1. timeout is GNU coreutils. It does not exist on a stock macOS install, so the shell answers command not found and every test file is graded failed.
2. ${command} is interpolated as a single argv token (same root cause as the static_analysis grader), so even with timeout present a multi-word test command would not run.

Repro

Fixture t.test.ts containing one trivially passing bun test case, graded via binary_tests:

"score": 0, "passed": false, "reasoning": "0/1 tests passed",
"results": [{"file":"t.test.ts","passed":false,"error":"bun: command not found: timeout\n"}]

while bun test t.test.ts in the same directory reports 0 fail / Ran 1 test.

Expected

1/1 tests passed.

Why it matters

Unlike the static_analysis sibling this one fails safe, but it makes the grader unusable on the platform LifeOS primarily targets: every passing suite is reported as a total failure.

Fix sketch

Drop the timeout binary in favour of Bun.spawn's own timeout option, and split the command into argv (or run it via sh -c) rather than interpolating it as one token.