fix(evaluation): restore term-based CAR scoring

[ShuxinLin]

Description

Reverts the CAR scorer change from #627 (f8805be). A CAR answer again passes only when it is a single-key object whose key matches the gold mode and whose value contains every required term extracted from the gold value. The score is the F1 over the mode key plus those terms.

Fix Details

• src/evaluation/scorers/static_json.py: restored byte-for-byte to its pre-f8805be state.
• src/evaluation/tests/test_static_json_scorer.py: the pre-f8805be CAR tests describe the groundtruth_eval.json metadata scoring added in 79d6356 (car_score = 0.6·mode + 0.4·coverage, optional terms, pass on any one required term). 8f1e538 took that out of the scorer two days later but left the tests, so 6 of them fail against this code. Replaced them with tests of what the code actually does: a missing term fails, every term is required, a wrong key still earns partial F1 from terms, the key is case-insensitive, a non-object answer scores 0, and the StaticJsonScorer wrapper reports pass/score.
• docs/static-json-evaluation.md: CAR section rewritten to match the restored behaviour. The old text still described the 79d6356 metadata scoring. No scenario in the aa_v1 suite ships a groundtruth_eval.json, so CAR terms come from groundtruth.txt either way.

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).

Every aa_v1 job graded before 9584e28 used this scorer. Re-scoring their saved CAR answers with it reproduces the recorded f1 exactly (0 mismatches across 223 trials in 5 jobs). Restoring it puts later runs back on the same scale.

CAR, 15 scenarios × 3 attempts, saved answers re-scored against the current gold:

Run	pass, mode-key only (#627)	pass, restored	mean f1, mode-key only	mean f1, restored
aws-claude-opus-5 high	48.9%	8.9%	0.489	0.362
azure-gpt-5.6-sol xhigh	44.4%	0.0%	0.444	0.224
gcp-gemini-3.6-flash high	51.1%	8.9%	0.511	0.311
MiniMax-M3	42.2%	6.7%	0.422	0.274
deepseek-v4.1-flash max	46.7%	2.2%	0.467	0.294
grok-4.7	53.3%	0.0%	0.533	0.262
glm-5.3	46.7%	0.0%	0.467	0.298

Only the gpt-5.6-sol xhigh job was graded with the #627 scorer, so its stored CAR rewards need re-verifying after this merges.

Related Issues

• Reverts the scoring change in fix(evaluation): score CAR scenarios on the mode key only #627

Verification Steps

1. uv run pytest src/evaluation -k "not integration": 111 passed.
2. uv run pytest src/ -k "not integration": 753 passed. 2 failures in src/observability/tests/test_file_exporter.py also fail on the base branch (missing google.protobuf locally).
3. Re-scored all saved aa_v1 CAR answers with the restored scorer and compared with the recorded verifier reports, as described above.

Checklist

• I have added tests that prove my fix is effective.
• My code follows the project's Ruff formatting and linting rules. (The two remaining ruff findings in static_json.py, SIM118 and ISC004, are in the restored code and also present on the base branch.)
• I have signed off my commits (DCO).