[yuecideng]

Description

Make standard RL rollouts, episode evaluation and saved environments reliable across ordinary tasks. This PR targets main and has no dependency on #981 or the assembly toolkit.

• Preserve owned pre-reset observations for completed rows and bootstrap only real timeouts; hard termination suppresses bootstrap.
• Infer terminal values only for eligible rows, with evaluation behavior and exact module-mode restoration, including exceptions. GRPO opts out of unused timeout-value inference.
• Evaluate PPO's clipped actor objective in log space without changing its finite objective or gradients.
• Add optional balanced per-environment episode quotas, completion/success denominators, Wilson intervals and scalar on_scored_episode records; retain environment-owned on_episode reports.
• Prepare scenes on owned configurations before World creation and export materialized resources through task-owned hooks. Saved Task Programs retain embodiment skill metadata and correctly bind physical/semantic references after configuration relocation. Synchronize packaging errors on all ranks before resume or training.
• Share device-reduced metric means between logging sinks, add optional final-target limits for relative joint actions, and group observation functors while preserving public and internally consumed import paths.
• Remove unused logging wrappers. Assembly-specific resource-path handling stays in the downstream task layer.

Stack

• Base: main at b629c7158b8180934c0990b71e12e1031ddac16e (Add compound rigid shapes and precision physics configuration #980 is already merged).
• Depends on: none of the assembly asset changes in Add portable mjfit assembly assets and precision demos #981.
• Followed by: Add temporal proprioceptive policy and auxiliary supervision #983. The task integration in Add native blind assembly tasks and reproducible policy replay #984 separately composes Add portable mjfit assembly assets and precision demos #981's assets with this runtime and Add temporal proprioceptive policy and auxiliary supervision #983's policy.

Type of change

• Bug fix and enhancement.

Validation

Isolated Python 3.12.14 / Torch 2.11 / DexSim 0.6.0rc1 environment, editable imports verified against this worktree:

• Affected Gym, actions, observations, runtime, evaluation, PPO, resume, CLI and metadata selection: 562 passed, 13 skipped. Includes a real two-process CPU Gloo regression: both ranks receive an export failure, then complete normal training/save/resume communication.
• GRPO/shared-rollout CPU, policy Task Program and context/API checker selection: 72 passed, 2 deselected.
• Existing Task Program bridge/policy/state/pour observation regressions: 154 passed, including the observation-facade compatibility paths.
• Actual wheel/source-free consumers and robot/example packaging: 5 passed.
• Native CartPole external-rollout integration under Xvfb/Hybrid: 1 passed (physics/data contract, not rendered-pixel quality).
• The independent runtime has no importable assembly toolkit or assembly benchmark package.
• Black 26.10.1 (black .), API coverage 4915/4915, context validation, changed Python headers/annotations, skill validation and whitespace checks passed.

Tests extend existing modules; compact/full terminal batches, hard termination, simultaneous termination/truncation, singleton BatchNorm inference and exception recovery, PPO/GRPO consumer selection, and file/inline Task Program snapshot reload close demonstrated gaps. Existing bridge coverage is reused for moved private observation helpers. Other mapped simulation/robot/randomization/data guidance remains accurate; RL lifecycle and observation ownership context was updated.

Focused collection comparison

Completed CPU collection latency, 32 steps without timeouts, FP32 [64,64] Tanh MLPs, one AMD EPYC 9K65 CPU thread, seed 42, five warmups and 15 alternating samples:

Environments	Previous #982 median	Current median
1	8.91 ms	7.91 ms
64	12.67 ms	10.79 ms
1024	56.92 ms	41.87 ms

Observation/action/reward/done/value tensors match exactly. Persistent rollout storage is unchanged. This measures collection only; GPU performance, CPU transient peak memory and end-to-end training throughput are unmeasured.

Limits

GPU cases are excluded from the CPU qualification. No rendered-pixel quality, NCCL or full task-training qualification is claimed. An initial expanded selection entered the native display initializer without a configured display and aborted; the isolated native external-rollout case subsequently passed with the proper Xvfb/Hybrid setup.

Checklist

• Run black . with Black 26.10.1.
• Validate affected behavior and review test necessity.
• Validate API coverage and review affected project context.
• Verify independence from Add portable mjfit assembly assets and precision demos #981 and retain downstream compatibility.

[han-xudong]

[P2] Preserve resume compatibility when freezing the saved Gym config

When a run starts with a component-based Gym config and is later resumed with trainer.gym_config pointing to that run's saved configs/gym.yaml, this expansion makes the snapshot incompatible with its own checkpoint. _configure_checkpoints() still stores the original declaration in _checkpoint_context, while _freeze_gym_config() replaces environment.component with fields such as env, physics, and max_episode_steps. checkpoint._validate() compares those raw contexts and raises Incompatible checkpoint configuration, even though both declarations resolve to the same environment.

A focused probe using an absolute component path (to exclude pre-existing relative-path issues) confirms that the base snapshot has no compatibility differences, whereas the new snapshot produces differences at metadata.context.gym.{env,environment,max_episode_steps,physics}. Both snapshots resolve to exactly the original environment. The added snapshot tests only reload the configuration; they do not resume the corresponding checkpoint.

Please use a consistent compatibility identity across the original configuration and exported snapshot, and cover resuming a checkpoint with its saved Gym config.