fix(m2-harness): calibration k>1 trials per item, distractor scoring on real dataset

- run-calibration.js: use k>1 repeated baseline trials per item for fractional
  difficulty in [0.40,0.70]; expose K_TRIALS env (default 3); remove unused
  buildConditionPrompt import; ceiling/floor flagged near 1.0/0.0
- run-distractor-scoring.js: wire LIMIT and full paths to real routing-cases +
  glm-distractor dataset with deterministic fired derivation; --fixture mode
  kept as supplementary unit test; LIMIT-sliced smokes yield n_scored>0 with
  real fpr/fnr/net_utility