feat(evaluation): add an fcc scorer for tolerance and set-F1 scoring from scenario_meta.json

[ShuxinLin]

What

Adds a deterministic fcc scorer (src/evaluation/scorers/fcc.py) and registers it next to fmea. A scenario selects it, and one of two match types, in its scenario_meta.json:

{"scoring_method": "fcc", "match": "numeric", "numeric_tolerance": 1}
{"scoring_method": "fcc", "match": "set_f1", "f1_pass_threshold": 0.8}

• numeric: the gold is a number, a percentage (trailing % stripped), or a list of numbers. The answer passes if round(|answer - gold|, 6) <= numeric_tolerance. For a list, the answer must be a list of the same length with every element within tolerance. A first line holding only the value is read directly, so numbers in an optional free-text line after it are ignored. Otherwise the whole answer goes through parse_structured_answer, and then the first line is tried as a fallback. Booleans never count as numbers. Score is 1.0 or 0.0.
• set_f1: gold and answer must be lists, and an answer that isn't a list scores F1 = 0. Distinct items are compared without regard to order. Scalars are compared through normalize_value, so case and whitespace don't matter. Objects are compared as whole objects, so a right wo_id with a wrong failure_code counts as one missing item and one extra item. F1 is 2·tp / (|gold| + |answer|). The answer passes if round(F1, 6) >= f1_pass_threshold, and the score is round(F1, 6).

Both comparisons are inclusive and use values rounded to 6 decimals, the same rule the fmea scorer uses since #646. If the metadata is bad (unknown match, a tolerance that is missing, negative or boolean, or a threshold that is not a number in [0, 1]) or the ground truth can't be parsed, the scenario fails with a rationale. The scorer never raises, so one bad file can't abort a run. details carries the gold and parsed values, the deltas or precision/recall/F1, the threshold and its operator, and the missing and extra items.

Scenarios that use it

match	setting	scenarios
numeric	numeric_tolerance: 1	303, 306, 307, 308, 309, 310, 311, 318
numeric	numeric_tolerance: 3.3	319, 323
set_f1	f1_pass_threshold: 0.8	304, 312, 316, 327

Every set_f1 scenario uses 0.8, the same threshold as the FMEA scenarios. These scenario_meta.json files live in the scenario-generation repo and are not committed yet. Until this scorer is merged, those files would make the evaluator raise unknown scorer 'fcc'.

Tests

• New src/evaluation/tests/test_fcc_scorer.py, 44 cases using synthetic values. It covers inclusive tolerance edges (including float error at the edge), percentages, element-wise lists and wrong lengths, the first-line rule and its fallback, F1 exactly at the threshold, order, case and duplicates, non-list answers, whole-object pair matching, invalid metadata or ground truth without raising, and registry lookup.
• python -m pytest src/evaluation/tests: 172 passed.

Integration check (not committed)

I loaded scenarios_data/ with the evaluator's loader. Exactly the 14 scenarios above resolve to scoring_method == "fcc", and each scenario's own ground truth passes as the answer (14/14). Below are pass counts on the 264 recorded runs for these scenarios, fcc against static_json. They are indicative only, because these runs predate a ground-truth correction.

scenario	match	runs	fcc pass	static_json pass
303	numeric	75	53	20
304	set_f1	9	7	0
306	numeric	10	4	0
307	numeric	10	8	7
308	numeric	10	4	3
309	numeric	10	5	1
310	numeric	8	0	0
311	numeric	10	7	1
312	set_f1	10	4	1
316	set_f1	10	7	0
318	numeric	74	45	16
319	numeric	10	8	1
323	numeric	9	6	0
327	set_f1	9	8	2
total		264	166	52