fix(terminal-bench): distinguish verifier infrastructure failures from zero scores #4571

Problem

Terminal-Bench 2.1 verifier infrastructure failures can become cached, unmasked zero rewards. This pre-existing behavior was raised during #3902 review and is separate from the sandbox lifecycle repair.

At e98eac14ac5708020472d88387d1be62634b9327, _verify catches verifier upload/setup/exec errors (the repair candidate retains this behavior in the extracted _verify_sandbox), then returns evaluation_completed=false, reward=0.0 without setting the common failure/masking fields. The Resources session caches this response. A repeated verify does not execute the failed operation again, and the resulting zero can be confused with a valid incorrect task result.

Reproduction and evidence

Independent local reproduction on the #3902 repair candidate uses the real verify → _verify → _verify_sandbox path and the canonical session replay, with only the sandbox provider fixture mocked:

1. Seed one Resources session.
2. Make _upload_folder raise RuntimeError for the verifier upload.
3. Call verify twice with the same request.
4. Observe both responses: evaluation_completed=false, reward=0.0, mask_sample=false, failure_kind=null, failure_reason=null.
5. The upload executes once, no reward is downloaded, and owner close still stops the sandbox once.

Source: verifier error path. This is a deterministic fixture reproduction, not an observed model-scoring failure or benchmark prevalence estimate.

Requested follow-up

• Select the benchmark contract for infrastructure failures: a typed no-result outcome (feat(core): define typed rollout outcomes and recovery #2135), or a completed masked result using the fields introduced by feat(core): define completed verifier-result failure semantics #2608.
• Preserve legitimate zero scores as unmasked task evidence.
• Make replay consistent with the chosen terminal/retryable outcome; do not rerun a potentially mutating verifier implicitly.
• Add regressions for upload/exec/reward-download failures versus valid zero reward, including session replay and reporting coverage.
• Verify aggregate reporting does not silently include unscored infrastructure attempts in the measured-score denominator.

No benchmark scoring change is included in the focused lifecycle repair. Owner: unassigned; benchmark/core outcome maintainers should select the contract before implementing it.