[0z5a]

The native backend rejected colocation, and its rollout worker requested the entire 1-CPU/1-GPU placement bundle. Removing the guard alone would leave the learner's additional 0.4 CPU/GPU reservation unplaceable and retain the wrong rollout device offset.

This adds an explicit resident GRPO/RLTT profile: one learner and one rollout engine share a bundle, with rollout requesting 0.5 CPU/GPU and reporting offset zero. The recipe disables both offload paths. Model state remains resident in separate workers; Ray fractions do not partition GPU memory. Dedicated placement, training math, publication barriers and held-out evaluation remain unchanged.

Stacked on #13. Author/committer and DCO: 0z5a.

Validation

580 distinct CPU tests pass; six CUDA tests skip and three known missing-Triton factory tests are deselected. This includes the existing numerical/update/recovery contracts, actual argument resolution, captured deployment/learner resource options, and family/algorithm/resume recipe cases. The initial one-case fixture failure is retained; the fixture omitted a field normally resolved by common validation. Ruff, formatting and source diff checks pass.

Scope	Parent #13	Resident candidate	Parent/current speed ratio
Native single-GPU resident profile	Rejected	CPU placement/argument checks pass	Undefined until actual runs
Matched official full RL	NOT_RUN	NOT_RUN	—
Whole-device peak memory	NOT_RUN	NOT_RUN	—
Held-out reward time-to-quality	NOT_RUN	NOT_RUN	—

Actual shared-device Ray/CUDA startup, capacity, complete training and fresh-worker recovery remain unqualified. A compatible existing environment and full resource handoff are required for those runs. Critic, offload, multi-rank learner, release-train and recovery are outside this initial resident profile; bare --colocate still selects offload defaults and is rejected.

Report: docs/native_colocation_results.md. Tested source hashes and raw receipts are retained in benchmarks/results/native-colocation-cpu/.