test(v7): mechanism ranking dataset, gates and feature exploration

* scripts/diagnose-v7-mechanism-ranking.js builds a deterministic population of
  UNSOLVED same-tier cards, a balanced sparse mirrored battle graph, a
  Bradley-Terry MechanismEmpiricalTheta, a mechanism-family split (whole
  victoryPath|resource families held out) and the spec's ranking metrics:
  Spearman, predicted/real spread ratio, pairwise direction accuracy for
  |delta| >= 1. Labels depend only on the generator, so they are cached by
  content hash and reused across model iterations.
* 48 cards / 432 edges / 2,592 battles: draw rate 1.9%, real mechanism spread
  p95-p5 = 6.264 theta, and every candidate model is blind to it (predictTheta
  rho -0.187, predictRealityTheta -0.065, BattlePower V4 -0.043). The barrierType
  ward-vs-shield lead (2.12 theta) was disproved by a controlled same-card swap
  (15 vs 14, 11 draws) - it is a PRNG-stream confound.
* Engine-accurate structural features reach best |rho| 0.453 (dpsToEhp) and
  held-out family Spearman 0.49-0.54 against the 0.85 bar, which per spec
  section 30 blocks the Solver/Full-Reality stage.

Author: NBL Agent