add invalid retry storm guard

[wang-tong0]

Summary

Adds an automatic guard for challenger models that repeatedly produce invalid evaluation attempts without writing those invalid attempts into sample_results.

What changed

• Executor workers now publish compact invalid-attempt summaries in worker_status_<env> under invalid_subjects.
• Invalid attempts are categorized broadly as model, stream, infra retryable, or env invalid.
• Scheduler checks for long-running invalid retry storms before the overlap gate and terminates the challenger only when the endpoint is still healthy and infra errors are not dominant.
• Added tests for executor invalid reporting, scheduler auto-termination, and the infra-retryable non-termination path.

Why

A model like UID247 can repeatedly emit malformed tool calls or broken structured output. Those attempts should not become valid zero-score samples, but without a separate signal the scheduler can keep retrying the same invalid work indefinitely.

Validation

• pytest -q tests/test_executor_worker.py tests/test_scheduler_flow.py tests/test_state_store.py tests/test_db_cli_sample_progress.py -q
• python -m compileall -q affine/src/executor affine/src/scheduler affine/src/scorer tests/test_executor_worker.py tests/test_scheduler_flow.py
• git diff --check