[0z5a]

ScaleRLT: I/O-Aware End-to-End Reinforcement Learning for Recurrent Looped Transformers.

Shared-prefix training retained the full prompt producer graph and each suffix wave through backward. Checkpoint the prompt producer and each complete suffix wave while preserving all loop/layer KV boundaries, accumulated prefix adjoints and all-loop readouts. Explicit token and label inputs reject response mutation before replay.

Stacked on #27. This first joint schedule supports a single whole-depth segment; layer/query subdivision remains outside this change.

Validation: 158 CPU passes, 4 CUDA skips and 12 unchanged strict BF16 xfails. Across 24 paired nonzero AdamW updates, maximum gradient error was zero and every trainable Adam moment matched the existing tolerances. The reproduced token-mutation failures and raw results are retained. Parent d6229d6 separately passed both real H20 MCore DDP/Adam cases; the new joint replay has not run on CUDA.

Result tables report saved-tensor observations and explicitly leave GPU speed, peak memory and online reward convergence unmeasured.