v3 Slice 4: Full causal domain - all 20 bundles x 5 variants + remaining scorers

[mark-allwyn]

Parent

#13

What to build

Grow the single B01 tracer into the full causal domain, exercising the remaining answer shapes end-to-end through the pipeline.

• The remaining part-scorers with their normalizers: edge_list (A->B, A->B, (A,B) -> canonical sorted pairs) and sequence_match, plus hardened set_match / numeric_tolerance / categorical (fractions, scientific notation, percent stripping, brace/bracket/delimiter handling, empty-set synonyms, verdict synonym classes).
• Generators for all 20 causal bundles x 5 variants, including the DAG transfer / transportability families that use the edge_list and sequence_match parts.
• Every generated item passes the dual-verification admission gate; every item is built only from refusal-neutral cover stories and passes a banned-vocabulary lint that rejects health/bio/security terms at generation time.
• Difficulty parameters biased so frontier models trend toward the 50-60% band (verified informally here; the canonical freeze happens later).
• The full set runs end-to-end on one live model.

Acceptance criteria

• All five part-scorers pass exact / tolerant-format / unparseable fixture cases
• Same seed produces byte-identical items across all bundles
• Solver vs simulation agreement holds on a large sample of random instances; disagreements reject items
• Banned-vocabulary lint is green over the entire generated set
• Full set runs end-to-end on one live model and produces scored results
• pytest green

Blocked by

• v3 Slice 2: Tracer bullet - one causal item end-to-end (generate, run, score, leaderboard) #15