[davek44]

PoissonMultinomialLoss.forward scaled y_true and y_pred by the position weights in place. The trainer computes the loss before updating its metrics (_train_batch_cov, _validate_cov). So the train/valid r, r² and spec it logs, and the stop_stat built from them, were computed on edge-downweighted tensors. This PR makes the multiply out of place. Loss values and gradients are unchanged.

Impact on logged metrics: with weight_range: 10, weight_exp: 8, the edge bins are scaled to 0.1. Down-weighting them inflates valid r by about 0.005 overall, and by up to about 0.015 for broad histone marks (e.g. H3K36me3, H3K4me1). hound_eval scores unweighted tensors, so its numbers never matched the trainer's for the same checkpoint and fold. After this change they do. Metrics logged by earlier poisson_mn runs are position-weighted.

Evidence:

• For a trained checkpoint, re-running the trainer's validation loop with this loss call reproduces the logged metrics.tsv to 1e-4 in every group. Without the side effect, it matches hound_eval to 1e-4.
• New test test_position_weights_leave_inputs_intact fails on main and passes here. src/tests/test_loss_weight.py: 34 passed.

🤖 Generated with Claude Code