[KnightStackSys]

Summary

• Add tunex ... texel for fitting integer HCE Eval* options against a JSONL dataset of labeled native FEN4 positions.
• Use the engine's static evaluation, not search scores; convert side-to-move values into Red+Yellow team perspective before minimizing Texel logistic mean squared error.
• Implement bounded integer coordinate descent, game-grouped train/validation separation, and select the best held-out weights.
• Add atomic checkpoints with dataset/config hashes, resume safety, dataset/result validation, example configuration and Lucide-icon documentation.
• Add regression tests covering labels, perspective, validation, mock-engine fitting, and resumption.

Dependency

Requires the companion IronPhoenix PR adding eval -> info string eval cp N. Use UseNNUE=false so HCE weights affect results. SpellWarden's evaluator has no runtime Eval* options, so this fit is intentionally IronPhoenix-only for now.

Validation

GitHub CI pending; no real dataset fit or Elo improvement is claimed. Confirm the chosen weights with independent paired SPRT.