Radar Studio 0.2 • 10 October 2026

Vitals continuity fix
- Vitals no longer disappear when ordinary tracking frames omit a vitals TLV.
- Every decoded frame in a USB chunk is processed for vitals, even when its last
  frame contains only tracking data.
- Last valid estimates and up to 150 actual received waveform samples are held
  separately for each subject. Sparse frames never replace them with zeros.
- Each subject shows the receipt time, source frame number and age of its vitals.
  Live vitals older than five seconds are labeled STALE, even if tracking frames
  are still arriving. Offline, replay and synthetic values are labeled separately.
- No interpolation or missing-sample estimation is performed. New acquisition
  sessions clear retained histories; switching presentation does not.
- Existing cards/canvases are reused instead of removed/recreated every frame.

Five seconds is a display freshness threshold, not a medical validity criterion
or claim about the firmware's measurement cadence. This fixes a confirmed UI
state bug; it does not establish that the hardware/USB stream never drops data.

3D view and track trajectories
- One interactive sensor-coordinate 3D view replaces the separate top/front
  plots. Drag to orbit, wheel/pinch or +/- to zoom; Top and Front remain presets.
- X/Y/Z axes, sensor origin, view bounds, and Reset help orient the scene.
- Optional per-track trajectories have 5/10/30/60-second arrival-time windows
  and a Clear control. Synthetic demo data also leaves visible trails.
- Trails split after a missing target, skipped frame numbers, a >2-second
  reception gap, or session reset. IDs are firmware track IDs, not verified
  people identities, and reuse without an observable gap cannot be identified.
- Trails cover only what this one radar reports in its view, not a whole-house
  map. Reception timestamps are not synchronized measurement timestamps;
  accelerated replay uses playback arrival time, not original capture timing.

Version 0.1 — initial prototype
- Dependency-free browser UI, separate CLI/data Web Serial connections, explicit
  configuration send with acknowledgements, point/track top and front views,
  experimental vitals, synthetic demonstration and local raw-byte replay.
- Project owner reported live operation on their Chromebook.
- Known sparse-vitals display limitation addressed in v0.2; no broad hardware
  validation or medical accuracy claim.

Public release packaging
- Added beginner launch/troubleshooting and developer handoff documentation.
- Vendor firmware/configs, private manifests and recordings omitted.
- At initial v0.1/v0.2 packaging, no license had been selected; current LICENSE and LICENSE_COVERAGE.md now apply as stated.

Version 0.5 — 10 October 2026 — experimental motion-signature prerelease
- Added original local sitting/standing/walking signature collection and a
  standardized nearest-centroid browser CPU baseline. Existing TI demo firmware
  still supplies points/tracks; no new on-board firmware is included.
- Added measured bracket-top reference height, signed antenna-center offset,
  tilt, tolerance and separate point/track coordinate interpretations; initial
  calibration is unset and must be verified. Motion-lab ROI/transform only.
- Added three-second non-overlapping human-labelled windows, separate training
  and held-out sessions, unknown/missing-data handling and local dataset I/O.
  Two training sessions per class and one held-out session per class are minimum
  gates, not an accuracy or generalization guarantee.
- Added optional bounded-memory complete received-UART capture, decoded frames
  and experiment events, independent of classifier acceptance. Explicit local
  export is not an upload and does not provide raw ADC or full micro-Doppler.
- Added MOTION_LAB.md with real-data capture, no-transition labels, leakage
  prevention, calibration, privacy and a clearly future custom-firmware roadmap.
- No real-data activity accuracy, medical validity or safety-critical claim.
  See VERIFICATION.txt for checks actually performed. Prior assets unchanged.
- Project LICENSE preserved. Captured-data license status remains pending;
  capture does not authorize dataset publication or participant-data sharing.
- Final v0.5 safeguards require measured null-default calibration, firmware/CFG
  setup identity, warning-free frames and matching profile/signature. Independent
  capture timelines, not renamed adjacent sessions, count toward split coverage.
- Manual sit-to-stand and movement-onset markers save unreviewed human click
  intervals in UART capture while stable labels pause. They are not automated
  event detection; starting movement is not necessarily walking. Camera/AI-label
  review is future work; current feature labels remain human-observed.

V0.6 (2026-10-10): local offline train/save/load/propose/review tools; separate activity/event models; immutable input files and separate overlays; capture-heldout metrics/unknown/coverage. Manual sit-down marker added. No pretrained model, upload dependency, original-firmware adapter or real-data accuracy claim. See OFFLINE_ML.md.
