refactor(reward): centralize post-score operations

[leviking98z-rgb]

Summary

• Add unirl.reward.ops as the single framework owner for reward-service dispatch, TensorRef materialization, scalar statistics, and lineage credit assignment.
• Route synchronous AR, diffusion, PE, unified-model, agentic, async, and multi-reward eval paths through that owner. Async reap keeps a dedicated attach-without-hydration operation so transfer remains deferred until a buffered sample is selected.
• Move propagate_rewards off the Sample transport type and into the reward layer.
• Extend the existing framework contract guard so trainers cannot grow direct score_and_attach or lineage-propagation paths again.

This is an ownership refactor, not a LOC-reduction PR: it is 270 insertions / 148 deletions (net +122), including 49 lines of static guard. It removes seven copies of reward hydration/statistics logic and prevents the wire type from owning reward policy.

This is stacked on refactor/bounded-recipe-composition (#30 in this fork), not on main.

Related Issue

N/A

Test Plan

• pre-commit run --all-files — passed, including recipe composition, architecture, framework, and experimental-boundary guards.
• python -m compileall -q unirl lint — passed.
• python lint/check_framework_contracts.py — passed; reports one reward-operation contract in addition to the existing framework contracts.
• python lint/check_architecture_boundaries.py — passed (476 files / 4,279 internal import edges).
• One-off CPU harness under torch==2.7.1+cpu exercised service dispatch, real-tensor materialization, component materialization, count/sum/mean statistics, and two-parent mean credit assignment — passed.
• Not run: GPU training, rollout, or remote reward-service smoke. Reward backends and score numerics are unchanged.

Compatibility / Risk

• Sample.propagate_rewards(...) is removed. Out-of-tree callers must use unirl.reward.ops.propagate_rewards(sample, ...).
• Trainer behavior is preserved: synchronous paths materialize at the same post-score point, and async paths still defer materialization until training consumes the selected batch.
• RewardOutcome.mean preserves the previous float32 mean and returns 0.0 when no reward is attached.
• No recipe, checkpoint, dataset, backend inference, or HTTP wire-format change.

Reviewer Notes

• Suggested review order: unirl/reward/ops.py, the PE/unified-model call sites, then check_reward_operations in the framework guard.
• Duplicate-work check: official feat(reward-service): add videohpsv3 scorer for T2V rollouts Tencent-Hunyuan/UniRL#270 adds a standalone video scorer and does not touch trainer-facing reward operations. Official feat(advantage): add GDPO per-reward-dimension decoupled group normalization Tencent-Hunyuan/UniRL#271 adds GDPO advantage policy and overlaps trainer/diffusion.py / types/sample.py, but not this responsibility; a textual rebase conflict is possible because both move policy out of trainer flow. No open PR was found that centralizes reward materialization and lineage credit.
• AI assistance was used. The complete diff and CPU/static validation evidence were reviewed before publication.

Checklist

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