[e-dobrowolska]

Standardize judge-failure recording across all LLM-as-judge benchmarks

Adds a shared abstraction (nemo_gym/judge.py) so a failed judge call (auth, rate-limit, timeout, HTTP, empty/malformed response) is recorded as a distinct outcome instead of being silently scored as a wrong answer:

• JudgeFailureMixin — standard per-sample fields judge_failed + judge_failure_reason (raw error), identical across benchmarks.
• run_judge() — wraps a judge call, catching failures instead of raising or defaulting to 0.
• judge_failure_metrics() — emits judge_failures (count), mean/judge_failed (rate), and two accuracy views: the usual one (failures = 0) and reward[judge_ok_only] (failures excluded), so a broken judge can't silently deflate scores.

Scope: applied to all 27 judge-based benchmarks (25 directly + physics_judge/math_with_autograder via inheritance), each with a wiring test.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[ananthsub]

Thank you for doing the original cross-benchmark judge audit and implementation. The shared helper and broad inventory informed merged #2113. This draft is now superseded by #2113 for the current helper and #3180 for migration away from reward-zero failure sentinels under the typed #2135/#2608 contracts. Closing this PR avoids two competing judge-failure designs while preserving its tests and server inventory as implementation evidence.