fix(evaluation): score CAR scenarios on the mode key only

[ShuxinLin]

Description

Score clarification-abstain-response (CAR) scenarios only on the top-level mode key of the answer (response, clarification or abstain). The text under the key no longer affects pass or score.

Fix Details

• src/evaluation/scorers/static_json.py, _evaluate_mode_json: an answer passes when it is a JSON object with exactly one top-level key and that key matches the gold mode (case-insensitive). score, f1 and strict_exact_match_accuracy are 1.0 on a pass and 0.0 otherwise.
• Required-term coverage (mode_required_terms, mode_matched_terms, mode_term_coverage) is still reported as a diagnostic, but no longer feeds pass, score or the precision/recall/F1 key counts. The per-key details now hold only the mode-key comparison.
• Harbor runs score with the same scorer through tests/test.sh, so Harbor rewards change too.
• Tests: six CAR tests were already failing on aafeedback_changes. They targeted the groundtruth_eval.json term scoring and car_score that were removed in 8f1e538. I replaced them with key-only tests: a correct key with no matching terms passes, a wrong key with matching terms fails, the key is case-insensitive, an answer that isn't an object fails, and the StaticJsonScorer wrapper gives 1.0/0.0.
• docs/static-json-evaluation.md: the CAR pass criteria and fields now match the code, and the docs say static_json does not read groundtruth_eval.json.

Impact on Benchmarking

• No change to baselines: This fix only improves stability/performance.
• Baseline change: This fix corrects a scoring error. (Please provide "Before vs. After" results).

CAR pass rates will go up for answers that pick the right mode but word it differently from the gold answer. Trade-off: an agent that always picks the same mode gets credit on every scenario expecting that mode. This PR has no before/after agent-run comparison.

Conflicts with origin/eval-scoring-relaxation (d1a668c), which rewrites the same function so that every required term is needed to pass.

Related Issues

• Fixes: #

Verification Steps

1. uv run pytest src/evaluation -q -k "not integration": 110 passed (6 failed before this change).
2. uv run pytest src/ -q -k "not integration": 752 passed, 2 failed in src/observability/tests/test_file_exporter.py. Those 2 also fail on the base branch without this change.
3. Manual: each gold answer of the 22 CAR tasks bundled under benchmarks/harbor/datasets/ passes when scored against itself.

Checklist

• I have added tests that prove my fix is effective.
• My code follows the project's Ruff formatting and linting rules. (No new findings; the 3 ruff reports in these files are already on the base branch.)
• I have signed off my commits (DCO).