[0z5a]

Depends on feature/huginn-looped-rl at 71860e55800229d68fbf48aed64a4a2688e8f1bd (upstream draft vllm-project#464). This fork PR contains only the incremental readout/objective work.

Add a response-only, fixed-depth readout for Ouro, Nanbeige and Huginn. It includes the prompt-to-response boundary, runs Huginn's coda for every supervised loop, and reconstructs vocabulary tiles during backward. A separate unclipped RLTT objective primitive makes reference policy, terminal sampled/full KL, reduction and credit gradients explicit. Existing dense provider calls retain their behavior; online trainer dispatch is still pending.

Test result

111 CPU tests pass: 50 numerical/objective oracles, 42 three-family readout comparisons, six provider regressions, and 13 existing short-update/publication/resume regressions. FP64 finite differences, FP32/BF16 readout derivatives, all trainable recurrent parameters, recomputation, empty responses and logical-batch partition invariance are covered. Ruff and diff checks pass. Three additional provider-factory checks fail at the existing MCore training import because Triton is unavailable; no packages were changed.

FP32, one CPU thread, three independent process starts; each warms both arms and runs three dense/streamed/streamed/dense blocks. These are geometric-mean readout-only forward/backward times on a shared macOS host. The prototype is slower in all measured cases; this is a correctness and retained-storage foundation.

Response / vocabulary	Objective	Dense (ms)	Streamed (ms)	Dense / streamed speed	Saved intermediate bytes before → after
64 / 1024	selected	0.316	0.765	0.414×	262,144 → 1,024
64 / 1024	selected + entropy + full KL	0.656	1.478	0.444×	786,432 → 1,024
256 / 4096	selected	5.186	7.179	0.722×	4,194,304 → 4,096
256 / 4096	selected + entropy + full KL	10.284	15.720	0.654×	12,582,912 → 4,096

Raw observations are in benchmarks/results/streamed-readout-cpu/; objective differences and gradient bounds are in docs/looped_readout_contract.md. Saved intermediate bytes exclude supplied inputs/parameters and do not measure peak memory or GPU memory. This is an RLTT objective reimplementation, not original-stack reproduction. Official-model GPU E2E, BF16 cross-engine scores, full-step acceleration and reward convergence remain unmeasured.