Summary

ADR-151 enrollment can accept an occupied room as the empty anchor — bench-reproduced on real ESP32-C6 HE20 captures and fixed here by replacing the per-frame presence statistic (median of |z| over subcarriers) with the per-frame p90 of |z|, re-deriving the gates from measured data, and excluding the unsanitised phase channel from motion.

The defect (measured, not theorized)

AnchorRecorder computes presence_z as the mean over frames of CalibrationDeviationScore::amplitude_z_median — the per-frame median of |z| across all active subcarriers (enrollment.rs, scored in wifi-densepose-signal ruvsense/calibration.rs::deviation).

The median-floor math: a person perturbs a minority of subcarriers strongly. For the majority of bins that stay statistically like the baseline, |z| ~ |N(0,1)|, whose median is 0.674. The median over all bins therefore floors near 0.674 and barely moves when someone enters the room. The default gates (empty_max_z = 1.0, min_presence_z = 1.5) sit in a band the statistic can rarely reach in either direction.

Measured on real captures (ESP32-C6 XIAO, HE20 2.4 GHz, 256-bin frames, node 8, replayed over UDP into the unmodified d0e27e65 CLI):

Both old values are below every gate — so on the clean upstream binary, enroll fed 100 % person-moving frames ACCEPTED the empty anchor (presence_z = 0.93). The p90 separates cleanly because the upper tail is exactly where the body's multipath perturbation lives.

Second finding — phase channel is noise on the CLI path: CalibrationRecorder's documented precondition is PhaseSanitizer + phase_align.rs, but the CLI UDP path feeds raw phase. The resulting baseline phase_dispersion ≈ 0.965 (uniformly random), so phase_drift_median ≈ π/2 on every frame and the |Δφ| motion term inflated still-room motion rates to ~60–70 % pure noise.

The fix

• CalibrationDeviationScore gains amplitude_z_p90 (numpy-convention linear interpolation; the median stays as a secondary/diagnostic field) and phase_usable (baseline median Von Mises dispersion ≤ PHASE_DISPERSION_USABLE_MAX = 0.5; the bench's unsanitised baseline measures 0.965, sanitized pipelines land ≤ 0.3 per CalibrationConfig::max_phase_variance).
• motion_flagged and the recorder's |Δφ| motion term only use phase when phase_usable.
• AnchorRecorder gates presence on the p90; gates re-derived from the bench bands with margin: EMPTY_MAX_Z = 1.6, MIN_PRESENCE_Z = 1.9 (constants, with the bench numbers cited in the doc comments).
• calibrate banner now prints z_p90 alongside z_med.

Bench proof (replay harness, same-epoch baseline)

Harness: parse the Pi's raw csi_raw bins, replay node-8 256-bin frames over UDP at 50 Hz into the rebuilt CLI; calibrate a 600-frame baseline from an empty segment, then run enroll (first anchor = empty) against held-out segments from the same capture window (same-epoch requirement — cross-epoch baselines drift; mean-shape correlation across epochs measured only ~0.31).

Separation 1.21 vs 2.54 against a gate of 1.6, consistent with the offline study bands.

Tests

• wifi-densepose-calibration: 43 unit tests green (suite updated for the p90 statistic; new regression tests: presence_uses_p90_not_median_floor, unusable_phase_does_not_create_motion) + full_loop integration green (sim now models the measured empty z-scale: real baselines integrate drift into the Welford variance, so runtime z ~ N(0, 0.82) — matching the measured empty p90 of 1.27–1.33 vs 1.645 for ideal i.i.d. noise).
• wifi-densepose-cli: 19 tests green. wifi-densepose-signal: all suites green.
• No new clippy warnings (per-file warning counts identical to d0e27e65).

🤖 Generated with claude-flow