Feat/bad syntax cancels eval

[ibraheem-abe]

• Adds two early termination checks during screener evaluations:
  • Threshold impossible — if the agent can't mathematically pass the threshold even if it passes every remaining run, skip the rest
  • Syntax penalty — if the agent produces a patch with invalid syntax (error code 1040), cancel all remaining runs and fail the agent out of screening
• Platform sends pass_threshold to the validator in the request-evaluation response
• Validator tracks results as runs complete using asyncio.wait(FIRST_COMPLETED) instead of asyncio.gather
• Cancelled runs are marked as skipped (new terminal status) via a new skip-evaluation-run endpoint
• evaluations_hydrated SQL view updated to treat skipped as terminal and detect syntax penalties
• handle_evaluation_if_finished transitions agents to failed_screening on syntax penalty instead of re-queuing
• Sandbox cleanup added to to tear down docker resources
• Only affects screeners — validators are unaffected (pass_threshold is None for valis)

Files changed

• validator/main.py — orchestrator updated, RunOutcome dataclass, cancellation logic
• api/endpoints/validator.py — pass_threshold population, skip endpoint, penalty handling
• api/endpoints/validator_models.py — new skip request/response models, pass_threshold field
• api/src/backend/postgres_schema.sql — updated view + skipped enum migration
• models/evaluation_run.py — skipped status added
• evaluator/sandbox/sandbox_manager.py — cleanup_sandbox() public method