refactor(rollout): seat colocated actor pools from the manager, not module state

[Rockdu]

What

• Remove the module-level placement-group state in miles/rollout/rm_hub/core.py (set_manager_placement_group / get_manager_placement_group and the set/get_reward_placement_group aliases) and the _encode_actors global in sft_rollout.py.
• RolloutManager now seats the colocated reward pool itself in __init__ (rm_hub.create_reward_pool(args, pg), keyed on --rm-type); pickscore_rm keeps its AsyncPickScorePool(args) lookup, which hits that singleton.
• AsyncRewardActorPool takes placement_group= explicitly. A colocated pool constructed without one raises immediately — this is also the fail-fast for mixed GPU reward types under --colocate-reward (previously the second pool's 0.05-GPU actor pended in Ray forever).
• call_rollout_fn passes placement_group= to rollout functions whose signature declares it (same pattern as evaluation= on custom_generate); sft_rollout.generate_rollout uses it to build a SftEncodePool singleton. Rollout functions without the kwarg are untouched.

Why

The reward pool is built lazily inside the rollout function, which has no handle on the manager, so the placement group was smuggled through a module global. The pool outlives every rollout call — its owner is the RolloutManager, whose __init__ already decides where the engines sit (init_rollout_engines), so it is the one object that knows which bundles carry a rollout claim. #207 was about to add a second global (_reward_slots) on top of this one; seating pools from the manager gives that PR a place to hang its bundle accounting without any module state.

GenerateState was deliberately not used as the owner: its lifecycle is one rollout call (reset() per call), while the pool is process-lifetime.

Validation

• pytest tests/fast/rollout/test_rollout_placement_group.py — 2 passed (kwarg is passed only to signatures that take it; an unseated colocated pool raises before creating actors).
• pre-commit run clean on the touched files.
• Not executed locally: the colocated pickscore path end-to-end. test_h3_t2va_grpo_2xGPU (H3 recipe, --colocate-reward --rm-type pickscore) covers it in this PR's CI; the three --colocate-reward recipes (H3, Cosmos3, Wan2.2 17-GPU) all use --rm-type pickscore, so create_reward_pool has a pool for each.

Files

• miles/rollout/rm_hub/core.py — drop the globals and aliases; placement_group= parameter; raise when colocated without seats.
• miles/rollout/rm_hub/__init__.py — create_reward_pool(args, placement_group).
• miles/rollout/rm_hub/pickscore.py — pass placement_group through.
• miles/ray/rollout.py — seat the reward pool in __init__; pass placement_group=self.pg to call_rollout_fn.
• miles/rollout/base_types.py — call_rollout_fn forwards placement_group when the signature takes it.
• miles/rollout/sft_rollout.py — SftEncodePool singleton replaces the _encode_actors global; generate_rollout declares placement_group=.
• docs/user-guide/customization.md — document the optional kwarg and who seats the reward pool.
• tests/fast/rollout/test_rollout_placement_group.py — new.

Follow-ups not in scope

• sft_rollout._scheduler_grid is still a module-level cache (a derived constant, not placement state).
• feat(reward): support HPSv2 and refine reward scheduling #207 should rebase onto this and drop _dispersal_order / ColocatedRewardSlots; with the pool seated by the manager, the remaining need is one actor per bundle, which the existing stride already gives.

🤖 Generated with Claude Code

https://claude.ai/code/session_019FxEmRysrrXzWUiQxxDR9k

[Rockdu]

Closing: this belongs on top of #207's branch, not as a standalone PR. Re-basing there.