Decouple run completion status from artifact scoring

[GTC2333]

Problem

BrainPilotBench currently marks a submission as submission_invalid and skips the grader whenever signals.completed is false, even if all expected artifacts were produced and passed submission verification.

Trigger scenario

A long-running task produces a complete model, inference script, and report, but the Agent is still auditing or wrapping up when the run ends because of a timeout, stream end, or another non-completed terminal state.

Impact

• Artifact quality cannot be distinguished from run-lifecycle failure.
• Complete and potentially valid artifacts receive no score.
• Harness and lifecycle regressions are harder to diagnose.
• Re-evaluation may require manually changing run state or repeating an expensive experiment.

Expected fix

Track artifact validity, Agent completion, and grader status separately.

For a submission with valid artifacts but an incomplete run:

• still execute the grader and save the artifact score;
• mark the run as timed out or incomplete;
• exclude it from the official leaderboard by default.

Submissions with missing or invalid artifacts should remain submission_invalid and should not run the grader.