[Tarzelf]

The Godot workstation can now record externally commanded PiPER block demonstrations, preserve interrupted attempts, and export explicitly reviewed camera/action rows for the existing CPU learner. This closes the capture-to-supervision connection; it does not claim a newly learned skill.

Stacked on #38 (feat/godot-workstation-v01). The older shoe workstation and existing demonstration validators remain unchanged.

• Adds a separate block recording UI and dvidia-demonstrate prepare/serve/inspect/recover commands: native 1 ms physics, 50 Hz control intervals and 25 Hz 320×240 RGB-D, with independent simulation and wall clocks.
• Uses encoder-only Cartesian control; preserves issued/applied targets, missing observations, source groups, native task evidence and source hashes. Scripted checks are explicitly labeled scripted, with no teacher calls or physical commands.
• Adds durable intent/outcome journals and crash-prefix reconciliation. A truncated journal remains unsealed and recoverable; it cannot silently become a training dataset.
• Adds dvidia-demonstration-export for reviewed segments with 269 causal inputs and seven direct issued-target labels, plus CI regressions and a fresh-environment install check.

Validation: 96 original Python test executions passed; 97 pass after the Windows inventory-path fix; Godot 7 coordinate + 29 recording assertions passed. The actual windowed scripted capture recorded 10 control intervals, six RGB-D pairs and 200 native ticks. Independent replay exactly matched all 10,200 native scalar values and the task scorer. A fresh installation outside the checkout inspected and exported all 10 selected rows. Peer review reproduced the durable-write interruption regression and confirmed its fix.

Windows CI caught backslash inventory paths in the initial revision. The fix serializes local artifact names as POSIX without weakening imported path checks. A fresh windowed capture again matched all 10,200 native values exactly; the base package was reinstalled and its export checked outside the checkout. Original evidence and the separate compatibility receipts are retained. Native rendering is still tested only on Mac.

The short motion check does not attempt placement; task success is false. Its 0.2 simulated seconds took 0.406678 wall seconds after handshake, so no sustained real-time qualification is claimed. Fresh installation checks cover the base CLI/export path; native graphics were tested in the existing simulation environment. No human demonstrations, model fits, hardware qualification or change to the previous 5/64 learned placement result. Source and compact evidence only; raw synthetic capture remains local. This draft is not a Hugging Face or research-site deployment.

Run guide · Evidence and timing scope