fix(hpo): sign-safe penalties, head_layer_size_length handling, honor random_state

[ChrisW09]

Fixes #420

Problems

1. A crashed trial can win the search. The exception path returned best_val_loss * 100. When the best loss is negative — routine for LSS models, whose val_loss is a mean NLL — that is an extremely good objective, so gp_minimize could return the crashing configuration as result.x.
2. Pruning threshold inverts for negative losses. best_epoch_val_loss * 1.5 sits below a negative best loss, so every subsequent trial exceeded the threshold at prune_epoch and was pruned, gutting the search.
3. head_layer_size_length corrupts head_layer_sizes. The trial loop special-cases the key first, but the final apply loop only checked key.startswith("head_layer_size_") — which the length key matches. The length (1–5) went through round_to_nearest_16 → 0, was prepended as a layer size, and the list was never truncated: [0, s1, ..., s5], a config no trial ever evaluated.
4. HPO ignored the estimator seed (random_state=42 hardcoded).
5. Sanity-check validation polluted val_losses, shifting epoch_val_loss_at() — the pruning baseline — by one epoch. Verified: a 2-epoch fit produced 3 entries.

Fix

Sign-safe arithmetic for both the penalty and the threshold, an explicit head_layer_size_length branch plus truncation in the apply loop, random_state plumbed into gp_minimize, and a trainer.sanity_checking guard in on_validation_epoch_end.

Tests

New tests/test_hpo_pruning.py: the penalty/threshold expressions stay worse than the incumbent across negative, zero and positive losses; the startswith collision and round_to_nearest_16(1..5) == 0 are pinned as the reason the length branch is needed; and a fit asserts len(val_losses) == 2 after 2 epochs.

🤖 Generated with Claude Code