Added her_sample_random_donors kernel to her_episode_kernel.cu.
Uses the existing per-sample LCG RNG state (already allocated in
GpuHer constructor, already used by Future strategy).
The per-step training hot path now has:
- Zero memcpy_htod (was 1 for HER random donors)
- Zero memcpy_dtoh
- Zero Vec/alloc
- Zero .clone()
- Zero cuStreamSynchronize
- Zero format!/String
Every CPU→GPU and GPU→CPU transfer has been eliminated.
The only remaining sync is the 1-step-lagged async readback
event check in replay_adam_and_readback (non-blocking in practice).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
e4f838d4e3