fix(evaluation): validate FMEA threshold, fix scorer-default drift

[ShuxinLin]

Description

Follow-up to #646, #648 and #649, from a cross-check of the recent eval changes against each other.

1. FMEA threshold validation. fmea.py took any int, float or bool from scenario_meta.json as f1_pass_threshold and silently used 0.80 for anything else (-1 passed every answer, 80 none, true meant 1.0, "0.8" was replaced by the default). The fcc and static_json scorers fail bad metadata with a rationale; fmea now does too.
2. Default scorer drift. fix(evaluation): let scenario_meta.json route scoring, and make the F… #646 made static_json the Evaluator/CLI default, but evaluation.runner.evaluate still defaulted to llm_judge.
3. Stale docs. docs/fmea-evaluation.md still said "strictly greater than 0.70". docs/evaluation.md, the Harbor README, QUICKSTART and the test.sh template still described llm_judge as the CLI default. docs/static-json-evaluation.md didn't mention relative_tolerance (feat(evaluation): add relative_tolerance to static_json scoring #649).

Fix Details

• fmea.py: _f1_pass_threshold returns the 0.80 default when the key is missing or null, and None for a bool, a non-number, NaN, or a value outside [0, 1]. The scenario then fails with error: "invalid_scenario_meta".
• runner.py: default_scoring_method is now static_json.
• Docs: inclusive per-scenario FMEA threshold; the CLI default; fmea/fcc rows in the scorer table; Harbor routing; the real cause of unknown scorer 'llm_judge'; a static_json "Numeric tolerance" section.

Impact on Benchmarking

• No change to baselines: All 13 FMEA scenario_meta.json files set a valid threshold (0.8 or 0.9). Re-scoring the 304 recorded FMEA trials raises no invalid_scenario_meta, so every verdict is unchanged. The runner.py default only applies to scenarios with no scoring_method when called through evaluation.runner.evaluate, which nothing in the repo calls outside tests.

Related Issues

• Follows fix(evaluation): let scenario_meta.json route scoring, and make the F… #646, feat(evaluation): add an fcc scorer for tolerance and set-F1 scoring from scenario_meta.json #648, feat(evaluation): add relative_tolerance to static_json scoring #649

Verification Steps

1. uv run pytest src/ -v -k "not integration": 853 passed. The 2 failures in src/observability/tests/test_file_exporter.py also fail on aafeedback_changes and are unrelated.
2. New tests in test_fmea_scorer.py: a scenario threshold overrides the default, a missing/null threshold uses 0.80, and true, -0.1, 1.5, 80, "0.8" and NaN fail with a rationale.

Checklist

• I have added tests that prove my fix is effective.
• My code follows the project's Ruff formatting and linting rules. (ruff check passes. fmea.py and its test were already off ruff format on the base branch, so the new lines follow the surrounding style instead of reformatting the file.)
• I have signed off my commits (DCO).