[0z5a]

Fixed-depth RLTT replay can now combine projection, MLP, normalization and response-head work across sequence tokens. CPU metadata resets each sequence's positions and preserves Huginn latent identity. The default stays serial; sdpa-reference batches non-attention work, while the explicit varlen branch calls Torch's causal packed attention and rejects unsupported device/dtype without fallback.

Incremental draft on #8. The 282-check regression passed, two CUDA checks were skipped, and the three known missing-Triton factory failures were deselected. A follow-up 41-case RLTT suite passes with both serial and packed modes, including fresh sampling, output-derived reward, nonzero AdamW updates, physical publication and exact fresh recovery; these test counts overlap. Actual Ray/CUDA lifecycle and official-model qualification remain pending.

Three independent CPU process starts measured the following serial/packed forward-backward ratios. The models are tiny (width 8, vocabulary 13, 16–30 tokens); attention remains segmented SDPA. B1 has no stable benefit. These are not GPU or full-RL speedups.

Tiny model	B1	B4	B16
Ouro	0.992×	2.709×	5.944×
Nanbeige	0.924×	2.757×	4.702×
Huginn	1.014×	2.405×	3.736×

The results report includes all 324 observations, process ranges, numerical tolerances and the retained failed FP32 gate/FP64 calibration. No peak-memory or reward-convergence claim is made.