[0z5a]

Per-layer checkpointing still retains recurrent boundary states and per-loop response inputs. Add opt-in loop segments with readout inside the checkpoint, physical-layer groups, and causal query chunks with full K/V-prefix gradients. Huginn coda remains outside the recurrent feedback state. A finite profile selector keeps workload, timing-scope and memory-metric contracts separate.

Stacked on #9 (perf/packed-recurrent-replay); one incremental commit. The default replay path is unchanged.

Validation: 354 CPU checks pass; six CUDA cases skip and three known missing-Triton factory checks are deselected. Full-parameter gradient comparisons, nonzero updates, and three Tiny native sampling/reward/Adam/publication/fresh-resume cycles pass with rematerialization.

Results and raw data include 864 same-tree observations and 36 unchanged-parent observations across 12 fresh processes. Selected plans reduce the observed forward saved-storage metric by 78.30–83.43%; this is not peak/GPU memory. Clean-parent ratios are 1.119× Ouro, 0.649× Nanbeige and 0.517× Huginn, with all intervals including 1. Slower schedules and Huginn storage regressions remain in the full table. Target CUDA, official-model full RL and reward convergence remain pending.