eval_runner marks submission as completed when results.json is missing

[rust-toml]

Summary

evaluate_submission() can mark a run/submission as completed even when no results.json exists.

In validator/src/eval_backend/services/eval_runner.py, _prepare_results() returns:

• metrics: {"results_missing": True}
• score: None

when results.json is missing.
But evaluate_submission() later unconditionally sets:

• run.status = "completed"
• submission.status = "completed"

This creates a false-success state.

Impact

A failed or incomplete evaluation can be reported as successful:

• API and PR reporting show completed while score is pending/null
• consumers cannot reliably distinguish true completion vs missing outputs
• evaluation integrity is weakened by silent success on missing artifacts

This pattern appears in merged PR timelines where results_missing: True was reported under completed status (e.g., PR #1, PR #2).

Proposed fix

Treat missing results as a failure (or explicit non-success terminal state), not success:

1. In evaluate_submission(), after _prepare_results(...), if metrics.get("results_missing") is True (or equivalent missing-output condition), set:
   • run.status = "failed" (or incomplete)
   • submission.status = "failed" (or aligned non-success status)
   • clear run.error/run.message stating results.json was not produced
2. Keep completed only for valid parsed outputs.
3. Preserve diagnosability by keeping results_missing in metrics.

Acceptance

• Missing results.json no longer yields completed
• API/PR reporting clearly indicates non-success status for missing outputs
• Existing success path for valid results remains unchanged
• Unit tests cover:
  • missing results -> non-completed terminal status
  • valid results -> completed
  • malformed/empty outputs -> explicit failure message