feat(evaluation): score CAR concepts from groundtruth_eval.json

[ShuxinLin]

Stacked on #630. Merge #630 first, then retarget this PR to aafeedback_changes.

Description

CAR scenarios score against the curated concepts in their groundtruth_eval.json instead of terms scraped from the gold sentence. This is a port of the CAR part of d1a668c (#617, closed unmerged). The numeric and time relaxations in that commit are not included.

• Concepts: each entry is {"concept", "any_of": [...]} or a plain string. A concept is covered when any of its surface forms appears as whole words in the answer, so paraphrases are accepted. Without the file, concepts are extracted from the gold value as before.
• Wrong mode: concepts are read only when the mode key is right, so a wrong-mode answer scores 0. Before this change, it earned partial credit for echoing terms.
• Pass: right mode + every required concept.
• Score: f1 is the F1 over the mode key plus the required concepts, (1 + k) / (1 + n), the same counting as every other static_json scenario. d1a668c replaced f1 with car_score. That made f1 >= 0.5 equivalent to "picked the right mode", so this PR keeps car_score as a reported field only.
• mode field: checked against the gold key, never used instead of it. A mismatch is reported as mode_spec_matches_gold (with an average in the report).
• Optional concepts: reported only, never scored.
• Wiring: StaticJsonScorer now passes scenario.evaluation_metadata to the scorer. The loader and the Harbor task generator already carried groundtruth_eval.json.

Type of Change

• New Benchmark Scenario (Industry/Asset type)
• Evaluation Metric / Scorer
• Agentic Orchestration Logic (ReAct, Plan-Execute, etc.)
• Infrastructure / Tooling Improvement

Impact

This takes effect only for scenarios whose folder holds a groundtruth_eval.json. The private suite's CAR files (scenarios 151–200) were restored unchanged from before their removal on 2026-10-07. Harbor tasks need regenerating so tests/scenarios/ carries them.

CAR results on the aa_v1 runs (15 scenarios × 3 attempts), with saved answers re-scored offline:

Run	right mode	strict pass: #630 → this PR	f1 ≥ 0.5: #630 → this PR	mean f1: #630 → this PR
aws-claude-opus-5 high	48.9%	8.9% → 26.7%	42.2% → 48.9%	0.362 → 0.426
azure-gpt-5.6-sol xhigh	44.4%	0.0% → 4.4%	17.8% → 37.8%	0.224 → 0.301
gcp-gemini-3.6-flash high	51.1%	8.9% → 20.0%	35.6% → 48.9%	0.311 → 0.412
MiniMax-M3	42.2%	6.7% → 8.9%	17.8% → 35.6%	0.274 → 0.306
deepseek-v4.1-flash max	46.7%	2.2% → 20.0%	24.4% → 44.4%	0.294 → 0.377
grok-4.7	53.3%	0.0% → 20.0%	26.7% → 51.1%	0.262 → 0.414
glm-5.3	46.7%	0.0% → 17.8%	28.9% → 44.4%	0.298 → 0.364

Related Issues

• Refs: Add scoring relaxations to the static_json scorer #617, fix(evaluation): restore term-based CAR scoring #630

Testing & Validation

• Unit Tests: uv run pytest src/evaluation -k "not integration" passes (118). uv run pytest src/ -k "not integration": 760 passed. The 2 failures in src/observability/tests/test_file_exporter.py also fail on the base branch (missing google.protobuf locally).
• Scenario Validation: loaded the 15 aa_v1 CAR scenario folders with their groundtruth_eval.json through load_scenarios and scored all 312 saved answers with StaticJsonScorer. Pass and score match d1a668c's verdict and exact_f1 on every answer.
• Data Integrity: no scenario content. The tests and docs use the synthetic "main unit" example.

Checklist

• My code follows the project's Ruff formatting and linting rules. (The remaining findings, SIM118 and ISC004 in static_json.py and UP017 in report.py, are on lines this PR does not touch.)
• I have performed a self-review of my code.
• I have updated the documentation (README or /docs) accordingly.
• I have signed off my commits (DCO).

[DhavalRepo18]

@ShuxinLin Please put the PR against aa_feedback and then reach out to them for the evaluation part only.