Verified on 2026-09-06 with Python 3.12.13, the committed uv.lock, and macOS Apple Silicon. Actual context: OpenGL 4.1 Metal - 90.5, renderer Apple M5 Max. Logical 1280×800 windows produced 2560×1600 Retina framebuffers. All automated windows remained hidden and unfocused.
Command from the repository root:
uv sync --frozen --project examples/python --python 3.12
bash scripts/verify-python.sh- 17 unit tests passed, with warnings treated as errors. Coverage includes fixed-offset rotating chains, shortest quaternion paths, parent-local reconstruction, batch/scalar interpolation equivalence, loop boundaries, pause/reset, exact JSON integers, finite values, unit quaternion norms, hierarchy/mesh validation, and all three canonical 1,299-frame clips. The malformed-value test contains 23 separate cases.
- Seven actual OpenGL captures passed: 8 seconds, 16 seconds, source overlay, A individually, a fractional frame at 14.3125 seconds, a narrow window, and a deliberately failed-pose diagnostic warning.
- The 8-second control check exercised all three solo selections, all-clips selection, source toggle, pause/reset, orbit/zoom, help hiding and resizing through the real viewer callback/render paths.
- All three 8-second robots produced actual mesh-only depth-tested sample counts of 56,242 / 47,174 / 43,810. Floor, trails, human overlay and HUD are excluded from these queries. Pixel readback independently contained 64,225 bright scene pixels.
- Scene-only changed-pixel ratios, excluding the HUD/footer: 4.3633% for 8→16 seconds and 0.1264% for source overlay on/off. These exceed the respective 0.2% and 0.03% gates.
- Two real CLI negative-input tests passed: non-integer schema and inconsistent fixed body offsets both exited nonzero with
G1_DATA_ERROR, no PNG, and no success marker. - Screenshots were visually inspected for complete bodies, readable white/dark mesh materials, source overlay alignment, unbroken mesh articulation, narrow framing, and the persistent kinematic/failed-pose labels. The bundled font's missing em-dash glyph is rendered explicitly in the permanent label.
The final local evidence set is under ignored artifacts/python/run-fhpodoeb/ (PNGs, per-run logs/JSON, and summary.json). Evidence directories are uniquely named on each verification run.
A bounded hidden render probe of 60 consecutive full three-robot frames completed in 0.880 seconds (68.18 frames/s) on this Mac, including the smoke occlusion checks. This is a local throughput measurement, not a cross-platform frame-rate guarantee. Batched hand-history interpolation uses the same fixed-offset math as single-pose interpolation.
Linux Mesa/Xvfb has a dedicated CI invocation; this macOS record does not certify a Linux result. Neither the unit tests nor graphics checks validate dynamics, self-collision, balance, contact, torque, policies, or real hardware. The display remains KINEMATIC REFERENCE — PHYSICS NOT VALIDATED.