fix(optimizer): log infill evaluations to TensorBoard in steady-state parallel runs

[bartzbeielstein]

Summary

Parallel runs (n_jobs != 1) logged only the initial design to TensorBoard: workers carry tb_writer=None (the writer cannot cross the dill/process boundary), and the parent steady-state result loop updated storage without ever writing per-eval hparams/scalars.

This adds parent-side per-eval logging in the batch_eval result handler:

• update_stats() refreshes counter (the TB step) which the steady-state loop otherwise never advances — kept inside the tb_writer guard so the no-TB path stays byte-identical.
• One add_hparams entry per evaluated point + one scalars write per batch, mirroring the sequential main-loop call site.
• Thread-safety: the result loop runs only in the parent main thread; eval workers never hold the writer and search threads only touch the surrogate under _surrogate_lock — single-writer, no lock needed.

Tests

• test_parallel_logs_infill_evals — asserts success_rate scalar advances past the initial-design step (fails on pre-fix code).
• test_parallel_no_tensorboard_regression — tensorboard_log=False parallel run: no writer, no runs dir.
• Full suite: 1745 passed; ruff clean.

🤖 Generated with Claude Code