[0z5a]

Huginn samples can have the same prompt and different recorded latents. The prefix learner now groups their actual latent identity and uses complete replay for singleton latent groups, preserving the original behavior scores, reference and logical loss. The actual Huginn provider enters through the opt-in FP32 MCore schedule and records shared/full-replay sample counts after backward finalization.

Stacked on #29. The related CPU regression passes 216 checks, skips 12 CUDA cases and retains 12 existing strict Ouro BF16 xfails. Twenty-four Huginn configurations compare two nonzero AdamW updates with complete packed replay, including all parameter gradients and Adam moments, a distinct reference, both reductions, ragged/empty responses, reordered microbatches and loss scales 1/8.

Recorded latent fixture	Shared samples	Paired updates	Max gradient relative L2	Max update difference	GPU speedup
Independent	0/5	16	7.175e-7	5.306e-7	NOT_RUN
Mixed	2/5	16	9.231e-7	2.941e-7	NOT_RUN
Shared	5/5	16	1.937e-6	1.460e-6	NOT_RUN

Detailed results and reproduction include the baseline comparison table, unchanged numeric gates, raw positive/negative logs and SHA manifests. Independent, mixed and shared arrangements are explicit fixed-trace fixtures; online sharing frequency is not established. Twelve actual CUDA DDP/Adam cases are prepared. Official Huginn CUDA, GPU memory/speed, scaled online RL and reward convergence remain pending; Huginn joint prefix rematerialization is unsupported.