feat: add could-not-evaluate status so non-gradable cases are excluded from scores

[pdebjyot]

What

Adds an EvaluationOutput.status field so an evaluator can signal that it could not produce a real verdict for a case, and excludes those non-gradable outputs from every score/pass aggregate. Implements the design in #346.

status is one of:

• "graded" (default) — score/test_pass are real verdicts and count toward aggregates.
• "could_not_evaluate" — the evaluator tried but couldn't grade the case (preconditions not met, missing data, harness error). Excluded from aggregates.
• "informational" — surfaces content for human review; never counts toward a numeric aggregate.

The default is "graded", so existing evaluators, consumers, and previously-saved reports are unaffected.

Why

Today, when an evaluator fails — a judge context-window overflow, missing data, unmet preconditions — the case is recorded as score=0, test_pass=False and that 0 is thrown into overall_score. A correct agent gets silently punished because the judge couldn't score the case, not because the agent was wrong. A capability failure of the harness is reported as a quality failure of the agent, with no way to tell the two apart.

How

• types/evaluation.py — add GRADED / COULD_NOT_EVALUATE / INFORMATIONAL constants (plain str, not an Enum, so new values can be added without a breaking change) and the status field on EvaluationOutput (Literal[...], default graded).
• evaluators/evaluator.py — _default_aggregator averages graded outputs only. When every output is non-graded there is nothing to score, so it returns a non-failure sentinel with the evaluator's own reason string.
• experiment.py — roll per-output statuses up to an evaluator-level status; tag both error-isolation paths (RetryError and generic Exception) as could_not_evaluate; compute overall_score from graded rows only.
• types/evaluation_report.py — flatten() excludes non-graded case rows from overall_score; rows with no status key default to graded so legacy reports load identically.

Raw per-case scores / test_passes lists are left untouched — only the aggregate changes, so anything reading individual rows is unaffected.

Behavior change (illustration)

A correct agent, one evaluator grading it 1.0, and one evaluator that raises (simulating a judge context-window overflow):

evaluator	score	status
GoodEvaluator	1.0	graded
OverflowEvaluator	0.0	could_not_evaluate

• Before: overall_score = 0.500 (naive average punishes the correct agent).
• After: overall_score = 1.000 (reflects only what was gradable).

Tests

18 tests across the three layers:

• Aggregator (_default_aggregator): graded-only average, could_not_evaluate/informational exclusion, all-skipped non-failure, empty-list unchanged.
• Report (flatten()): non-graded row excluded, missing-status defaults to graded, all-non-graded → 0.
• EvaluationOutput.status default + settable.
• Experiment layer: error-isolation path tags could_not_evaluate; mixed-status run excludes the non-graded rows from overall_score.

Full suite: 1753 passed (baseline + net-new tests, no regressions). ruff check / ruff format --check clean.

Closes #346