feat: prepare HandUMI research preview

[alvax64]

Classification

GitHub/source research preview only. No release, PyPI publication, documentation deployment, safety qualification, or scientific validation is authorized. TEST-001 has not passed.

Included

• Crash-safe staged capture, periodic privacy-safe manifests, disk/ENOSPC handling, bounded workers, stage profiling, recovery, checksums, and software soak tooling.
• Explicit compact dataset compatibility and pre-compact rejection policy with controller-only/body/recovery/corrupt golden fixtures.
• TEST-001 software-only transform, synchronization-event, metric, dropout, participant bootstrap, and report groundwork with synthetic evidence.
• Source/native dependency and robot-asset inventory gates; R1 Lite remains excluded.
• Whole-tree Ruff 0.15.22, formatter, and basedpyright 1.39.9 immutable CI gates.
• Recorder startup owns loopback-only Rerun through cleanup, rolls back partially opened cameras, requests MJPG, and requires simultaneous fresh camera streams before dataset creation.
• Local trajectory replay reads only required Parquet columns and directly decodes embedded image bytes; remote and true video-backed layouts retain the LeRobot fallback.

Verification on 866b592

• Full local matrix: 472 pytest tests with 37 unchanged upstream deprecation warnings; lock, compileall, formatter, Ruff, basedpyright, diff checks, and strict Sphinx all pass.
• Exact-head CI/package run 29956309419 and documentation run 29956311386 pass.
• Exact 297-frame no-video replay improved from 3m14s to 1.839s; camera-enabled replay completes in 3.856s with one workspace stream.
• Camera-enabled headless RRD: 10,813,297 bytes, SHA-256 3fdc28872a206bdb086f75a777ec788777a22acc5681ef4306c79da69c35681c.
• 30-minute software-only/headless soak on ancestor 82792e0: 1,800.101 s, 18,000 rows, 162,000 synthetic camera items, 9.9994 rows/s, zero drops/failures/timeouts, maximum queue depth 1, RSS 24,051,712 to 25,038,848 bytes, FDs 5 to 5.
• 30 Hz/three-camera software-only profile: 1,800 rows / 5,400 camera frames in 60.169 s, 29.916 rows/s, zero drops/failures/timeouts.

Connected diagnostics (not qualification evidence)

A worn-Quest three-camera/Rerun run completed 197 rows over 19.7 seconds and retained 3,085 native packet-v2 sidecar rows. A later workspace-camera diagnostic completed 297 rows over 29.7 seconds with 4,046 native packets, 23 valid canonical body joints per row, no viewer failures, and 11/11 verified file checksums. Body CoM/contact/support remained invalid because no fitted profile was supplied.

Offline quality rejected both short diagnostics. The workspace-camera run exceeded the 60 ms tracking synchronization tolerance on 3/297 rows (1.0101% versus the 1% threshold; 199.923 ms maximum). It used an explicit three-second controller recovery window and is visual diagnostic evidence only.

The right-wrist camera repeatedly disconnected and re-enumerated at the USB layer. The simultaneous startup gate now rejects this before dataset creation, but stable direct/powered USB hardware remains required before repeating the supported three-camera and 30-minute gates.

Still blocked

Legal/asset permission closure, calibrated physical Quest/camera/Feetech/robot gates, stable right-wrist camera power/cabling, 30-minute hardware soak, release signing/ownership authority, and TEST-001 laboratory execution remain external. Software-only or short visual evidence is not Quest lifecycle, camera endurance, thermal, radio, robot, anatomical-accuracy, physical-safety, or scientific-validation evidence.

Do not merge or publish solely from these results.