[Daniel-rmc]

The local workspace can now run robosuite and four RoboCasa kitchen tasks, expose fixed/tool-frame six-axis manual controls, and execute explicit text primitives through Jev. The workspace groups manual controls, model execution, and scene settings while keeping Stop available.

The integration also fixes renderer cleanup on reset, sensor-aware action schemas, request snapshot ordering, unstable manual-control menus, and reuse of verified LIBERO/VLA runtimes between worktrees. Optional simulator dependencies remain isolated, and policy observations retain the sensor-only boundary with separate task evaluation.

Jev uses its own server-memory TypeSafe credential and records choices, probabilities, source observations, and request counts. Its released mode is explicit text/proprioception control; the independent experimental visual-caption probe and planned joint-policy routing remain development work.

Validation on the candidate source:

• 426 tests and 136 subtests passed with real LIBERO, robosuite, RoboCasa, and all three local checkpoint checks enabled. Jev simulation tests use explicit synthetic provider responses; no live cloud model was called.
• Clean main checkout: 364 passed, 62 optional checks skipped, 136 subtests passed.
• Jev log checks explicitly read UTF-8 on Windows; 47 focused checks passed after the test-only encoding fix.
• All 7 JavaScript checks passed; CI now runs them on Linux, macOS, and Windows.
• Offline wheel build passed, packaged UI files match source byte-for-byte, and the wheel completed a real Panda rendering smoke test.
• Private maintainer records, credentials, runtimes, model weights, and raw experiments are excluded from the commits and package.