The two normalization branches of MixtureSampler disagreed on all-zero
input: temperature == 1.0 crashed with ZeroDivisionError at the divide,
temperature != 1.0 silently degraded to uniform sampling via the
max(w, 1e-12) clamp. Both are reachable from the annealing phase path
in scripts/train.py, where TrainingPhase.dataset_weights validates >= 0
(not > 0) to intentionally allow dropping individual datasets.

A user who zeroes out every dataset in a phase transition would get a
mid-run crash on one branch and silent training-quality drift on the
other. Add a single _validate_weights() helper that rejects empty,
negative, or all-zero weight lists with a clear ValueError, and call
it from both __init__ and update_weights. The TrainingPhase schema-
level validator stays permissive because it can't know the effective
mixture at config-parse time; sampler-level is where the hard gate
belongs.