[0z5a]

Repeated prompt groups currently recompute the same actor prefix in every suffix microbatch. This adds a fixed-depth gradient reference that runs prefix forward once, accumulates every loop/layer KV adjoint and direct first-response readout adjoint, then runs prefix backward once before the optimizer step. Ouro, Nanbeige and same-latent Huginn preserve the complete shared-parameter gradient; Huginn keeps coda separate from recurrent feedback and rejects a different latent identity.

Incremental stack on #10 (5f7ce7fa). This is an explicit serial-suffix reference API, not a default MCore/Ray schedule. The ordinary three-phase prefix technique is attributed to its existing work. Packed suffix execution, distinct-latent Huginn prelude sharing, joint rematerialization, distributed integration and target-GPU/full-RL validation remain pending.

Validation: 402 CPU checks pass, six CUDA checks skip, and three known missing-Triton factory checks are deselected. Thirty-six added full-gradient/update configurations have maximum relative L2 error 1.819e-7; separate FP64 attention and all Q/K/V gradients agree within 1e-12. Original overly strict FP64-weight failures and mixed-precision diagnostics are retained: native normalization still reduces in FP32.

The result report includes all speed-comparison tables and raw data. All 648 same-tree and 108 untouched-parent observations match loss, full gradients and nonzero SGD updates. The serial-suffix reference is slower: clean-parent P24/D4/G32 ratios are 0.404× Ouro, 0.326× Nanbeige and 0.471× Huginn. The parent batches suffix projections; this reference currently does not. No GPU memory saving, end-to-end RL improvement or reward convergence is claimed. All 688 parent blobs and provider import paths are checked; full parent gradient/update oracle vectors are retained.