fix(evaluation): score CAR on groundtruth_eval.json concept recall

[ShuxinLin]

Description

CAR scenarios ship groundtruth_eval.json again, but static_json never read it: the loader put it on Scenario.evaluation_metadata and the scorer ignored it. This PR scores CAR answers against its required_terms.

Fix Details

When a CAR scenario's groundtruth_eval.json has required_terms, the score is recall:

1. Mode gate. The answer gets recall 0 unless it is a JSON object with exactly one top-level key and that key equals the mode key of groundtruth.txt. A multi-key answer counts as a mismatch.
2. Concept coverage. Otherwise recall is the fraction of required concepts covered. A concept is covered when any one of its any_of forms, or a form's regular plural, appears in the value under the key. Matching is case-insensitive and on whole words.
3. Pass. The scenario passes at recall >= 0.5 (CAR_PASS_THRESHOLD).

score and the Harbor reward are that recall. precision and f1 are null for these answers. A free-text value gives no count of what the model asserted, so precision could only repeat recall, and an F1 built from them would be recall under another name.

Example with two required concepts, asset (asset / equipment) and time window (time window / date range):

Model answer	Recall	Passed
{"clarification": "Which equipment, over what date range?"}	1.0	yes
{"clarification": "Which assets?"}	0.5	yes
{"clarification": "Can you say more?"}	0.0	no
{"abstain": "Which asset and time window?"}	0.0	no

Other details:

• Plurals: a form also matches an added s, an es after s/x/z/ch/sh, or ies for a consonant + y (asset → assets, switch → switches, battery → batteries). For multi-word forms the plural goes on the last word. Irregular plurals still have to be listed.
• Metadata fields: mode and optional_terms are not scored. Bare-string required_terms entries are accepted as one-form concepts.
• Structured values: term matching read a structured value (an object or list under the key) through json.dumps. That kept upper-case letters out of the [a-z0-9] filter and turned a newline into \n, gluing an n onto the next word. It now joins the structure's keys and scalars and lowercases them.
• Reports: the stale car_score aggregate is removed from report.py.
• Unchanged: CAR scenarios without required_terms keep the extracted-term F1 and strict pass rule, with exact matching (no plurals). Non-CAR scoring is unchanged.

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).

Pooled re-score of 515 saved CAR answers (11 models), each through the loader and StaticJsonScorer:

Scorer	Pass rate	Mean score
Before (aafeedback_changes: extracted-term F1, strict pass)	3.7%	0.290
After (concept recall, pass >= 0.5)	39%	0.360
Reference: answers with the correct mode key	41%	n/a

Earlier CAR results are on a different scale and should be re-scored from saved answers.

Known limits of term matching: nothing penalises padding, so an answer that lists every form gets full recall. Negation is not detected.

Related Issues

• Follows fix(evaluation): restore term-based CAR scoring #630

Verification Steps

1. uv run pytest src/evaluation -q: 126 passed. In the full uv run pytest src/ -v -k "not integration", the only failures are 2 in src/observability/tests/test_file_exporter.py, which fail the same way on the base commit.
2. Re-scored the saved answers above with the base and branch scorers. On the branch, every CAR result came from the concept path (f1 is null).
3. Ran uv run evaluate on one CAR scenario: the report has f1_avg: null and recall as the score.

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; 2 pre-existing ones are on untouched lines.)
• I have signed off my commits (DCO).