[KnightStackSys]

Summary

• Add dependency-free discrete Tree-structured Parzen Estimator (TPE) Bayesian optimizer for integer UCI spin parameters.
• Use a random startup phase followed by smoothed good/bad density models, selecting candidates via a log likelihood ratio and avoiding duplicates.
• Evaluate each candidate through paired engine matches against fixed baseline options, using existing IronPhoenix or SpellWarden match runners.
• Record W/D/L, smoothed objective, parameter vector, and deterministic RNG state; save atomically and reject incompatible resume configs.
• Add tunex -c configs/bayesian.example.json bayesian [--iterations N], documentation, and tests.

Verification

Awaiting GitHub CI. No claim of measured Elo gains; validate selected settings with independent SPRT.

Notes

No Optuna or additional dependencies; TPE is a Bayesian optimization method rather than a Gaussian-process implementation. Boolean tuning remains a separate A/B test.