[pppppppppper]

Summary

Adds RTC asynchronous execution across EmbodiRun's simulation and physical-control paths.

• provides sync, naive, and rtc runtime modes with explicit overlap and modeled-delay settings
• emits the versioned metadata.async.rtc contract to EmbodiInfer and the SGLang π0.5 integration
• validates SGLang RTC endpoint capabilities before execution
• supports RTC in LIBERO and SO101 control, including camera and serial recovery needed by the real-robot loop
• keeps optional RTC numerical dependencies lazy so minimal host/recipe installations still start

The existing synchronous path remains the default when runtime_options is empty.

Companion inference PR: BUAA-CI-LAB/EmbodiInfer#2.

End-to-end LIBERO success rate

libero_spatial, 10 tasks × 5 trials, modeled delay Δ = 4 steps, identical seeds, and the standard LIBERO checkpoint:

Mode	Inference behavior	Trials	Success rate	Mean environment steps
sync	synchronous upper bound	50	1.000	101.9
naive async	asynchronous, no compensation	50	0.620	325.2
RTC	asynchronous with prefix guidance	50	0.960	113.4

RTC recovers 34 percentage points over the naive asynchronous baseline and finishes within 4 percentage points of synchronous execution.

Verification

ruff check and ruff format --check pass for the changed source and tests.

PYTHONPATH=src pytest tests/test_async_algorithms.py tests/test_async_simulation_runtime.py tests/test_sglang_async_wire.py tests/test_sglang_http.py tests/test_control_services.py -q

Result: 85 passed, 21 skipped in the non-SGLang environment.

PYTHONPATH=src:integrations/sglang_pi05/src pytest tests/test_sglang_http.py tests/test_sglang_async_wire.py -q

Result: 33 passed with SGLang 0.5.18 and Torch 2.13.

The previously failing shared-device/runtime compatibility case now passes.

Historical real-robot benchmark recordings remain local and are intentionally excluded from this PR.