feat(evaluation): add relative_tolerance to static_json scoring

[ShuxinLin]

Summary

static_json passes only on a strict exact match. Its delta-1 and range checks are reported but never feed the pass decision. This lets a scenario opt in to a relative tolerance on numeric values through scenario_meta.json, following the match vocabulary the fcc scorer uses:

{"scoring_method": "static_json", "match": "numeric", "relative_tolerance": 0.1}

• Every flattened gold value must be present, and an extra key fails the answer.
• A number passes when |answer - gold| <= relative_tolerance x |gold|. The check is inclusive and uses values rounded to 6 decimals, as in fcc and fmea.
• Any other value, such as an asset id, must match exactly. Keys stay positional, so a list must keep the gold order.
• The score is F1 over accepted keys. The existing static_json details are kept, and the per-value results are added under tolerance_values.
• A match other than numeric, or a missing, negative or non-numeric relative_tolerance, fails the scenario with an invalid_scenario_meta rationale instead of raising.
• Scenarios without a match key are scored exactly as before.

Why

health-409 and health-410 ask for five per-unit failure rates to one decimal. The key comes from the expert failure-mode tag in the excavator dataset, which the agent has to rebuild from the job text. Under exact matching, a reader who agrees with the expert on 98% of jobs still fails (D 122.4 against 125.2), so the scenarios couldn't be passed by a correct method.

The companion scenario change (in the scenario repo) adds relative_tolerance: 0.1 to both scenarios. It also states the rate window in the question: "measured from each unit's first to its last breakdown".

Effect on recorded runs

I regraded every recorded 409/410 answer (143) with this branch, loading the scenario files through load_scenarios:

Answer	409	410
Recorded runs (every job counted, 2–3x the key)	0/104 pass	0/39 pass
Expert labels, stated window	pass	pass
Expert labels, n−1 convention	pass	pass
Expert labels plus jump starts, or plus tooth/lip replacements	fail	fail
5% of jobs misread at random	88% pass	88% pass

No recorded pass/fail result changes.

Tests

uv run pytest src/evaluation -k "not integration" gives 186 passed. The new tests cover:

• inclusive boundaries
• the every-job answer failing
• order and asset ids held exact
• an extra key failing
• a fenced answer
• five invalid metadata shapes
• strict matching when match is absent