feat(B3/G4): health-scaled Expected SARSA temperature — breaks collapse attractor
When health=1 (healthy): tau unchanged → sharp softmax → near-argmax target (deterministic).
When health=0 (collapsed): tau scales 6× → wide softmax → stochastic sampling breaks Q-collapse attractor.
- c51_loss_kernel.cu: add get_learning_health() helper, isv_signals as last param, tau_base → tau * factor
- gpu_dqn_trainer.rs: add last_sarsa_tau_factor field + init, launch_c51_loss → &mut self, host-side mirror, append isv_signals_dev_ptr arg
- fused_training.rs: add last_sarsa_tau_factor() accessor
- training_loop.rs: propagate SARSA tau factor from fused ctx for HEALTH_DIAG logging
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
e07c82976e