Code Graders and the Verdict for an Experiment #1004

Parent

#977

What to build

After its Trials, an Experiment scores each Trial with code Graders wherever a check is mechanical:

• a Trial whose repository changed is void;
• a review that left no reports fails;
• each Eval task's own mechanical checks.

It then computes its Verdict with #999's module, and writes the Scores and the Verdict into its results. The Verdict also goes into its Command record's outcome: a fail or an inconclusive Verdict puts the Experiment in Needs you, and a pass puts it in Done. --json ends with one line holding the Verdict and its interval. Each Arm's minutes are shown beside it.

Acceptance criteria

• An integration test gives a fixture suite whose mechanical checks pass for one Arm and fail for the other, and gets the expected Verdict.
• A failed Experiment shows in Needs you, and a passed one in Done.
• --json ends with the Verdict line.
• thirdshift experiments <id> shows each Arm's minutes.