Flaky test: terminal_bench_4 test_verifier_boundary wall-clock bound #5434

resources_servers/terminal_bench_4/tests/test_verifier_boundary.py failed once in CI on a wall-clock assertion and passed when the job was re-run on the same commit.

Where it failed: line 406, in the check that a slow verifier preparation stays within its budget:

assert 0.18 <= elapsed < 0.28
E       assert 0.8193796259984083 < 0.28

The test's fake sandbox sleeps between 0.12 s and 0.5 s per step and then asserts a tight elapsed-time window, so on a loaded runner the measured time can exceed the upper bound without any change in behaviour.

Run: "Unit tests / Server suite (shard 6)" on commit 445e407 (PR #5343). The first attempt failed; the re-run of the failed jobs passed. No commit in that PR touches terminal_bench_4.

Possible fix: assert the ordering and the budget enforcement rather than an absolute wall-clock window (for example, that the verifier stopped waiting at the budget and returned the expected reward), or widen the upper bound to tolerate scheduler delays.

Similar to #5429 and #5433.