Record Godot demonstrations and export reviewed camera-action data

.github/workflows/ci.yml

@@ -61,6 +61,8 @@ jobs:
           dvidia-training-benchmark --help
           dvidia-skillspace-coverage --help
           dvidia-episode-tape --help
+          dvidia-demonstrate --help
+          dvidia-demonstration-export --help
           dvidia-reaction --help
           dvidia-motor-memory --help
           dvidia-observe --help
@@ -80,6 +82,8 @@ jobs:
           dvidia-training-doctor --json
       - name: Verify pinned learning caches
         run: python -m unittest tests.test_camera_action_cache tests.test_camera_cache_cli tests.test_camera_action_export tests.test_camera_action_model -v
+      - name: Verify operator capture, recovery and reviewed exports without physics
+        run: python -m unittest tests.test_operator_capture tests.test_operator_export tests.test_godot_demonstration_bridge -v
       - name: Verify exact-reference and initial-head contracts without physics
         run: |
           python -B -m unittest tests.test_camera_action_reference_export -v

README.md

@@ -6,6 +6,15 @@
 
 [Recorded demo](https://huggingface.co/spaces/Dvidia/dvidia-training) · [Synthetic datasets](https://huggingface.co/datasets/Dvidia/dvidia-training-examples) · [Pilot models](https://huggingface.co/Dvidia/dvidia-training-pilot)
 
+**[Record paired demonstrations in Godot](docs/godot-demonstration-v02.md).**
+Control the simulated PiPER's position and gripper, record synchronized native
+RGB-D and issued actions, then export explicitly reviewed segments in Terminal.
+Interrupted attempts and failed outcomes remain inspectable. A scripted windowed
+check reproduced all **10,200 native scalar values exactly** in independent replay.
+This is a capture/export prototype; human demonstrations, new policy training and
+real-time teleoperation qualification remain ahead.
+[Evidence and timing limits](benchmarks/godot-demonstration-v02/README.md).
+
 **[Verified local learning cache](docs/camera-action-export-cache-v01.md).**
 The **0.1.3.dev5 preview** adds `dvidia-train cache` to reuse verified camera/action
 features with caller-pinned data, split and source identities. On one 1,000-sample

benchmarks/godot-demonstration-v02/README.md

@@ -0,0 +1,89 @@
+# Godot demonstration capture: integration evidence
+
+This is a scripted recording/export check, not a manipulation benchmark or a
+new learned skill. The final recording contains ten 50 Hz command intervals,
+six 320x240 RGB-D pairs, and 200 native 1 ms ticks. It does not attempt a complete
+placement; the independent task scorer correctly returns failure, with no
+critical contact ticks. No human demonstrations or model fits were performed.
+
+The original evidence below is pinned to commit `b8b96cd`. Windows CI then
+identified a local inventory serialization bug: backslash paths were correctly
+rejected by the portable recording validator. The follow-up fix uses POSIX
+paths for locally generated inventory entries, retaining strict validation of
+imported references. `portability-tests.json` records **97 passing test executions**
+including Windows/POSIX path cases. `portability-checks.json` records a fresh
+windowed capture, exact replay of another 10,200 native values and a reinstall/
+export check after the fix. This second short run took **0.482349 wall seconds
+for 0.2 simulated seconds**; it is a compatibility check, not a speedup claim.
+Native rendering remains tested on Mac; cross-platform data tests do not
+qualify Windows rendering. Original receipts are retained unchanged.
+
+## Results
+
+| Check | Result |
+|---|---|
+| Python regression tests, including existing capture/export contracts | 96 test executions passed |
+| Godot coordinate and recording-contract assertions | 7 + 29 passed |
+| Actual windowed Godot: Begin, ten steps, Finish | Passed; zero client errors |
+| Independent native replay | All 10,200 scalar values exactly matched; maximum difference 0 |
+| Independently reconstructed task score | Matched; task unsuccessful |
+| Fresh installation outside checkout | Both CLI entry points work; installed source hashes match |
+| Reviewed direct-action export and reload | 10 rows x 269 causal features, 10 x 7 issued target labels |
+| Learner inputs | Existing feature contract; no teacher calls or native object/contact truth |
+
+Fault-injection tests cover interrupted intent/outcome writes, fsync before
+in-memory journal cursor updates, and truncated trailing records. Finalization
+reconstructs the committed prefix before sealing. A truncated journal remains
+unsealed and recoverable. Pause/resume/reset/disconnect were also exercised over
+real localhost TCP on an earlier source revision; those retained development
+captures are not presented as evidence pinned to the final recorder source.
+Their final-source behavior remains covered by the Python regressions above.
+
+## Timing scope
+
+Final windowed capture: **0.2 simulated seconds in 0.406678 wall seconds** after
+the initial handshake. The server reports 201.570 ms aggregate command handling,
+including controller, camera/perception, file writing and finalization. Twelve
+RPC samples include Begin and Finish; their p95 is 53.432 ms. This small sample
+does not establish a tail-latency bound, sustained throughput or 50 Hz real-time
+operation. Physics, wall-clock delivery and GUI frame rate are different metrics.
+
+Host: Apple M5, 10 logical processors, macOS 26.5.2 arm64, Python 3.12.9, NumPy 2.5.3,
+MuJoCo 3.15.0, Godot 4.7.2. The standalone viewer uses the compatibility OpenGL
+renderer. No GPU-training speed, RAM minimum or physical reaction latency was
+measured. Godot headless assertions passed with sandbox log/certificate-access
+warnings; the actual windowed run completed without client errors.
+
+Earlier development windowed receipts measured 0.511076 and 0.595010 seconds
+for the same short simulation duration, on earlier recorder/UI revisions. They
+remain local alongside the initial graphics-sandbox refusal and a development
+capture rejected after the source-identity contract changed. Do not infer a
+controlled speedup from these development runs.
+
+## Evidence and reproduction
+
+- `tests.json` pins tested source and reports the regression executions.
+- `windowed-client.json` retains the scripted request/state trace and client
+  source hashes. Its local output directory is replaced with a portable label;
+  other values match the original receipt.
+- `native-replay.json` records independently executed native array comparisons.
+- `capture-summary.json` retains the independent task score, source identities
+  and exact original capture inventory. The synthetic pixel/native capture is
+  retained locally; it is not bundled in this small source/evidence publication.
+- `fresh-install.json` records installed source hashes and exported array shapes.
+- `workstation.png` is the real final window screenshot, not a design mockup.
+
+Follow [the capture guide](../../docs/godot-demonstration-v02.md). To reproduce
+the short engineering check, prepare with `--duration 0.2`, serve with
+`--source-kind simulation_scripted_input --max-episodes 1`, then run Godot with
+`--smoke-steps 10 --smoke-output /absolute/fresh/receipt.json` and an optional
+`--screenshot /absolute/fresh/workstation.png`. Run the native replay script on
+the resulting episode. Use fresh directories; failed attempts are preserved.
+
+MIT applies to the added DVIDIA source. Upstream arm notices remain under
+`third_party/robot-models` and in prepared scene exports. MuJoCo is an external
+Apache-2.0 dependency. No credentials, real footage or model weights are included.
+
+Next: collect actual operator demonstrations, review useful recovery segments,
+and register a separate learning comparison on unseen starts. The existing
+5/64 learned placement result and its unqualified status remain unchanged.