feat(hslm): Phase 3 — PBT hyperparameter search

[gHashTag]

Summary

Phase 3: Population-Based Training for automated hyperparameter optimization.

New file

• src/b2t/pbt_search.zig — 253 LOC

PBT Algorithm

1. Initialize population of 5 members with random hyperparameters
2. Train each member for N steps
3. Evaluate fitness (negative loss)
4. Exploit: Replace bottom 40% with copies of top performers
5. Explore: Mutate hyperparameters (lr, label_smoothing, tau, weight_decay)
6. Repeat until convergence

Hyperparameter search space

• LR, batch_size, label_smoothing, tau_init, weight_decay, grad_clip
• Crossover: uniform selection from two parents
• Mutation: Gaussian perturbation with configurable strength

Tests (5)

• Population init, fitness tracking, exploit/explore, crossover, mutation

Closes #295