fix: isolate per-miner exceptions in OSS scoring loop

[plind-junior]

Summary

Wrap the per-miner evaluate_miners_pull_requests call in gittensor/validator/oss_contributions/reward.py in try/except Exception so a single miner's uncaught exception cannot abort the whole scoring round.

Without this, any unhandled exception (e.g. RecursionError from #978, ValueError from #893, future schema-drift KeyError / AttributeError) propagates through oss_contributions → forward → concurrent_forward → the while True at neurons/base/validator.py:144-169. The except Exception there logs once and run() returns. Because run() is the target of a daemon thread (base/validator.py:179), the scoring thread dies silently while the main thread keeps printing Validator running | uid X | ... every 30s — no further forward passes, no weight updates, until manual restart. After restart the validator re-fetches the same data and re-crashes on the same miner.

The crashed miner now gets a MinerEvaluation(uid=uid, hotkey=hotkey, failed_reason=...), so downstream skip-on-failure paths (MinerEvaluationCache.store_or_use_cached_evaluation, mirror_scan, repository.upsert_miner_evaluation) already treat the miner as non-scoring this round without any new per-site awareness.

This is a safety net, not a substitute for fixing each specific exception class at source. It brings the OSS path to parity with issue_discovery/mirror_scan.py:157-162, which already isolates per-miner failures.

Related Issues

Closes #995

Related (independent vectors that fire this chain):

• walk_node in tree-sitter scoring can hit Python's recursion limit on deeply nested ASTs #978 — RecursionError from tree_sitter_utils.walk_node on AST depth >1000
• [Bug] MirrorClient malformed 2xx JSON bypasses mirror failure handling and can stop validator rounds #893 — raw ValueError from MirrorClient._get response.json() not classified as MirrorRequestError

Type of Change

• Bug fix (defense-in-depth)

Testing

• Tests added
• Manually verified loop continues past a forced exception

New tests/validator/oss_contributions/test_reward.py:

• test_per_miner_exception_does_not_abort_round — boom on UID 3 (raises RecursionError); asserts UIDs 1/2/4/5 are still scored, UID 3 carries failed_reason, get_rewards returns a normal (rewards, evaluations, cached_uids, penalized_uids) tuple.
• test_failed_reason_records_exception_class_and_message — asserts the substituted MinerEvaluation.failed_reason includes both the exception class name and message, so post-hoc DB inspection can identify the vector that crashed a miner.

Local verification:

pytest tests/validator/oss_contributions/  # 124 passed
ruff check gittensor/validator/oss_contributions/reward.py tests/validator/oss_contributions/test_reward.py  # clean
ruff format --check gittensor/validator/oss_contributions/reward.py tests/validator/oss_contributions/test_reward.py  # clean

Checklist

• Code follows project style guidelines
• Self-review completed
• Tests added
• ruff check / ruff format --check clean
• Catches Exception (not BaseException) so KeyboardInterrupt / SystemExit still propagate
• No behavior change on the happy path — the wrap only adds a fallback

[anderdc]

neurons/validator.py:200-206 already detects scoring-thread death via thread.is_alive() and exits the process within 30s for supervisor restart — the "silent freeze until manual restart" premise doesn't hold. The OSS path is asymmetric with mirror_scan only because mirror_scan catches a specific MirrorRequestError, not bare Exception; this PR is strictly broader and would mask deterministic crashes (#978, #893) that should be fixed at source. Closing.