fix(m5-decisive-rerun): rerun 4 debugging skills on FROZEN 224-item decisive split, fix scorecard verdicts

- run-swe.js: add SWE_DATASET_PATH env var support (backward-compat, defaults to
  150-item external/swebench.jsonl); add dataset_path/n/provenance fields to
  result JSONs for dataset misuse detection
- Rerun five-whys-plus (+0.9pp, p=0.724), occams-razor (0pp, p=0.683),
  kepner-tregoe (-1.8pp, p=0.221), map-territory (+0.5pp, p=1.0) on n=224
  decisive split (EVAL_RUN=m5-primary-decisive, CONC=4, isolation ON)
- Provenance tag 'exploratory: surface-mismatch' written into map-territory
  and kepner-tregoe result JSONs (fixes blocker 2)
- Scorecard: occams-razor verdict corrected from DIRECTIONAL-NOT-REPLICATED
  to NO-LIFT (primary that fails >=5pp+p<0.05 is NO-LIFT; fixes blocker 3)
- Decisive split commit predates all 4 result files (VAL-DATASET-009)