refactor(rollout): split lifecycle and receiver capabilities

[leviking98z-rgb]

Summary

Refactors rollout engines around explicit, composable capability surfaces instead of a fat base class with unsupported NotImplementedError methods.

• Move the shared offload/health behavior into DefaultRolloutMemoryLifecycle and require every rollout engine to declare MEMORY_LIFECYCLE.
• Define dependency-light structural protocols for tensor, NCCL, IPC, LoRA, checkpoint, and memory-lifecycle consumers.
• Remove optional weight-receiver stubs from BaseRolloutEngine; unsupported transports are now absent and rejected by capability planning.
• Share tensor/NCCL/LoRA delegation across vLLM-Omni, SGLang AR, and SGLang diffusion while preserving engine-specific IPC, copy, checksum, and target_modules behavior.
• Validate capability-to-method correspondence both at execution-plan resolution and in the stdlib-only framework guard.

The change removes 429 lines of duplicated base stubs and engine forwards. The protocols, reusable implementations, docs, and persistent guard make this PR net +58 lines; the goal here is an enforceable extension boundary rather than a standalone LOC reduction.

Related Issue

N/A

Test Plan

• SKIP=no-commit-to-branch pre-commit run --all-files --show-diff-on-failure — passed, including recipe targets, architecture boundaries, and framework contracts.
• python -m compileall -q unirl lint/check_framework_contracts.py — passed.
• python lint/check_framework_contracts.py — passed: 7 rollout engines, 6 weight syncs, 19 model pipelines, 2 train backends, 2 wire types, 4 loop programs, and 13 entrypoints conform.
• One-off stdlib-only harness (not committed, per repository policy) — passed; it exercised tensor/NCCL delegation, AR target_modules suppression, weight-version increments, LoRA dirty-state forwarding, and rejection of incomplete declared method surfaces.
• GPU rollout/training smoke — not run; this environment has no torch/Ray/Hydra stack or accelerator allocation. No generation, numerical, or transport implementation is changed.

Compatibility / Risk

• No recipe, checkpoint, or wire-format migration is required for in-repo components.
• This intentionally removes unsupported receiver methods from BaseRolloutEngine. Out-of-tree code that blindly calls an unsupported transport now gets normal attribute absence instead of a delayed NotImplementedError; it should query declared capabilities.
• New rollout engines must declare MEMORY_LIFECYCLE and structurally implement every receiver capability they advertise.
• The highest-risk area is inherited receiver dispatch. The shared mixin preserves the previous signatures, version increments, and per-engine target_modules behavior, and the one-off harness covers those differences.

Reviewer Notes

• This is stacked on fork Draft PR refactor(trainer): separate lifecycle and loop programs #25 (refactor/trainer-loop-programs) and should be reviewed as the next architecture slice.
• AI-assisted (Codex); the full diff and listed checks were reviewed in-session.
• Duplicate-work check: upstream refactor(rollout): engine-level sync/async contracts for async rollout Tencent-Hunyuan/UniRL#287 owns sync/async generation and quiesce contracts. This PR is a complementary memory/weight-receiver follow-up and will need a mechanical rebase if refactor(rollout): engine-level sync/async contracts for async rollout Tencent-Hunyuan/UniRL#287 lands first. Upstream feat(weight-sync): add SGLang checkpoint-engine IPC sync Tencent-Hunyuan/UniRL#225/feat(weight-sync): support SGLang TP/EP full-weight sync Tencent-Hunyuan/UniRL#226/feat(rollout): support SGLang TP/EP LoRA sync Tencent-Hunyuan/UniRL#227 add SGLang checkpoint-engine, TP/EP full-weight, and TP/EP LoRA transports; if any land first, their new receiver methods should be represented by the corresponding capability during rebase. Upstream feat(rollout): per-lane dynamic dispatch for AR rollout (sync + async) Tencent-Hunyuan/UniRL#289 changes AR scheduling and does not duplicate this contract extraction.

Checklist

• I reviewed the changed code and removed unrelated/generated artifacts.
• I updated tests, docs, and configs where needed, or explained why not.