test(tasks): gate cube physical recovery

[yuecideng]

Description

Stack

• Layer: 14/14
• Base: feat/production-parallel-safety
• Depends on: feat(expert-program): validate parallel joint segments #482
• Followed by: Top of stack
  Close the current physical acceptance loop for declarative Expert Programs. The UR5 cube preset uses 100 real motion samples while retaining the 0.08-rad tracking gate and bounded replanning; a slow GPU gate proves all three lazy Pick/Place/settle/validator cycles complete through contact dynamics.

A second slow gate injects only a bounded controller-ready gripper-open command through ordinary env.step() calls during Place. It proves physical held-object loss, core-owned symbolic invalidation, RECOVERY_REQUIRED, a real re-acquisition Pick, Place retry, and completion of the remaining program. The fault wrapper never writes object pose, velocity, constraint, attachment, or task state.

This layer also documents the placement and parallel APIs and refreshes the deterministic rollout report.

Refs #471
Refs #474

Type of change

• Enhancement (non-breaking change which improves existing functionality)
• Documentation update

Screenshots

Not applicable.

Validation

• Relevant cross-stack regression — 1215 passed, 2 skipped, 8 deselected
• pytest -q tests/benchmark/expert_program/test_demo_success_cube_sim.py tests/benchmark/expert_program/test_cube_physical_recovery_sim.py -m 'slow and gpu and requires_sim' --run-gpu — 2 passed in 186.81s
• pytest -q tests/gym/envs/tasks/test_multi_segments_cube_pick_place.py tests/gym/envs/tasks/test_open_drawer.py tests/scripts/tools/test_expert_program_rollout_report.py — 19 passed, 1 deselected
• python scripts/tools/expert_program_rollout_report.py --check — up to date
• make html — succeeded; Sphinx reported 698 repository-wide warnings
• Black 26.3.1 check on changed Python files and git diff --check

Checklist

• Changed Python files pass Black 26.3.1.
• Public API, usage, rollout, and design documentation is updated.
• Unit, integration, and real-simulation tests prove the behavior.
• No dependency update is required.

[greptile-apps]

Greptile Summary

This PR strengthens the repeated-cube physical acceptance coverage and updates the corresponding task preset and rollout documentation.

• Increases the UR5 cube preset to 100 motion samples while preserving tracking and recovery bounds.
• Adds slow GPU simulation gates for three-cycle completion and physical-loss recovery.
• Refreshes rollout metrics, capability documentation, and public API references.

Confidence Score: 5/5

The PR appears safe to merge because no concrete blocking or independently actionable non-blocking defect remains.

The production change is a bounded task-policy calibration, and the added simulation gates exercise the intended completion and recovery paths without exposing a supported failure mechanism.

Important Files Changed

Filename	Overview
embodichain_tasks/embodichain_tasks/multi_segments/cube_pick_place.py	Configures 100 motion samples for the cube preset while retaining existing tracking and bounded-recovery policies.
tests/benchmark/expert_program/test_cube_physical_recovery_sim.py	Adds an isolated simulation gate that injects controller-level gripper opening and verifies physical loss, invalidation, reacquisition, and retry.
tests/benchmark/expert_program/test_demo_success_cube_sim.py	Adds a slow GPU regression gate validating all three physical Pick/Place cycles and their effect, settling, and validator metadata.
scripts/tools/expert_program_rollout_report.py	Refreshes the deterministic capability and integration matrices to represent newly validated coverage.
docs/design/expert_program_rollout_report.md	Updates generated rollout metrics and physical-acceptance status consistently with the reporting tool.

Sequence Diagram

sequenceDiagram
    participant Test as Recovery gate
    participant Env as env.step()
    participant Physics as Contact dynamics
    participant Runtime as Expert Program runtime
    Test->>Env: Replace bounded Place commands with gripper-open targets
    Env->>Physics: Apply ordinary controller commands
    Physics-->>Runtime: Held-object relation contradicted
    Runtime->>Runtime: Invalidate symbolic held state
    Runtime->>Env: Execute recovery Pick
    Env-->>Runtime: Re-acquisition verified
    Runtime->>Env: Retry Place and remaining program
    Env-->>Test: Completed three-segment episode

Loading

_{Reviews (1): Last reviewed commit: "test(tasks): gate cube physical recovery" | Re-trigger Greptile}