…churn

The Live Demo "Human Pose Detection" rendered a black canvas with stacked,
scrambled labels ("tudo preto e tudo em cima de tudo").

Two root causes, both in the pose pipeline:

1. Keypoint confidence was structurally dropped at the tracker boundary.
   `PoseTrack::new`/`update_keypoints`/`create_track` only accepted [[f32;3]]
   positions, so `KeypointState.confidence` kept its 0.0 init value forever.
   The server emitted all 17 keypoints with confidence 0, which (a) made the
   renderer skip every joint (guard `conf > 0.1`) → black canvas, and (b)
   collapsed the bbox to the degenerate 0.6x1.0 fallback → every "Conf: 90%"
   label piled onto one point. Fixed by threading per-keypoint confidence
   through create/update; centroid-filled slots stay 0.0 (unobserved).

2. The Kalman tracker is tuned for world-space metres (measurement_noise
   0.08 m, Mahalanobis gate 9.0) but was fed image pixels, making the effective
   association gate ~1.5 px. A single person's natural ±20 px jitter never
   matched its own track, so every frame spawned a fresh track and stale ones
   lingered as confirmed ghost skeletons ("Persons: 4/5" for one real person).
   Fixed by converting px<->m at the tracker boundary (PIXELS_PER_METER=100).

UI: pose-renderer keypoint-confidence thresholds made inclusive (>=), since the
server clamps confidence to [0.1, 1.0] and joints at the floor were dropped.

Tests: confidence round-trip, multi-person separation, and single-person
no-churn regression tests added. Verified end-to-end against the real binary:
stable "Persons: 1" over 427/427 frames, one clean human skeleton.

Co-Authored-By: claude-flow <ruv@ruv.net>
Claude-Session: https://claude.ai/code/session_01BztNyDLaawF9UUCJ1wrwTz

Independent feature in wifi-densepose-pointcloud (camera depth backend);
no overlap with the pose-tracker fixes on this branch. Clean auto-merge.

Co-Authored-By: claude-flow <ruv@ruv.net>
Claude-Session: https://claude.ai/code/session_01BztNyDLaawF9UUCJ1wrwTz

Most recent of the automated submodule-update branches; the other 7 point at
the same ruvector SHA, so only this one is merged.

Co-Authored-By: claude-flow <ruv@ruv.net>
Claude-Session: https://claude.ai/code/session_01BztNyDLaawF9UUCJ1wrwTz

…ything

Records the two pose-pipeline fixes (keypoint confidence round-trip, px<->m
tracker scaling) under [Unreleased] Fixed, and the merged ADR-271 Depth Anything
Rust camera-depth backend under Added.

Co-Authored-By: claude-flow <ruv@ruv.net>
Claude-Session: https://claude.ai/code/session_01BztNyDLaawF9UUCJ1wrwTz