fix(train): wire wifi-densepose-signal into the pipeline; correct MODEL_CARD env-sensor claim (#536)

Addresses three findings from the 2026-05-11 training-pipeline audit:

#1/#2 — `wifi-densepose-signal` was a phantom dependency of `wifi-densepose-train`
(listed in Cargo.toml, never imported), and vitals/CSI signal features were
absent from the pipeline. New module `wifi_densepose_train::signal_features`:
`extract_signal_features(&Array4<f32>, &Array4<f32>) -> Array1<f32>` (and the
convenience method `CsiSample::signal_features()`) runs a windowed observation's
centre frame through `wifi_densepose_signal::features::FeatureExtractor`,
producing a fixed-length (FEATURE_LEN=12) amplitude / phase-coherence / PSD
feature vector — the hook for a future vitals / multi-task supervision head
(breathing- and heart-rate-band power are read off the PSD summary). The vector
is produced on demand and is not yet fed back into the loss; wiring it as a
training target is the documented follow-up. `wifi-densepose-signal` is now an
actually-used dependency. 5 new tests (2 unit in signal_features.rs, 3
integration in tests/test_dataset.rs); existing wifi-densepose-train tests
unchanged and green.

#3 — `docs/huggingface/MODEL_CARD.md` presented PIR/BME280 environmental-sensor
weak-label fine-tuning as a current capability; there is no env-sensor
ingestion in the training pipeline. Marked that path as planned/not-implemented
in the training-steps list and the data-provenance section.

(#5 — README's "92.9% PCK@20" overclaim — fixed separately in PR #535.)

CHANGELOG updated.

CHANGELOG.md

@@ -18,6 +18,13 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
   PowerPlatePulse training-pipeline audit (2026-05-11); 6 remaining audit findings
   tracked in the PR.
 
+### Added
+- **`wifi-densepose-train`: `signal_features` module — wires `wifi-densepose-signal` into the training pipeline.** `wifi-densepose-signal` was previously a phantom dependency of `wifi-densepose-train` (listed in `Cargo.toml`, never imported). New `wifi_densepose_train::signal_features::extract_signal_features` (and `CsiSample::signal_features()`) run a windowed CSI observation's centre frame through `wifi_densepose_signal::features::FeatureExtractor`, producing a fixed-length (`FEATURE_LEN = 12`) amplitude/phase/PSD feature vector — the hook for a future vitals / multi-task supervision head (breathing- and heart-rate-band power are read off the PSD summary). The vector is produced on demand and not yet fed back into the loss. Surfaced by the 2026-05-11 training-pipeline audit (findings #1 "vitals features absent from training" and #2 "`wifi-densepose-signal` ghost dep").
+
+### Fixed
+- **HuggingFace `MODEL_CARD.md`: marked the PIR/BME280 environmental-sensor ground-truth path as planned, not implemented** (training-pipeline audit finding #3) — the card presented PIR/BME280 weak-label fine-tuning as a current capability; there is no env-sensor ingestion in the training pipeline today.
+- **README: corrected the camera-supervised pose-accuracy claim** (audit finding #5; see PR #535) — "92.9% PCK@20" → the ADR-079 target (35%+; proxy baseline 35.3%), noting P7/P8/P9 are pending.
+
 ### Added
 - **`nvsim` crate — deterministic NV-diamond magnetometer pipeline simulator** (ADR-089) —
   New standalone leaf crate at `v2/crates/nvsim` modeling a forward-only

docs/huggingface/MODEL_CARD.md

@@ -168,14 +168,14 @@ The training process works like this:
 1. **Collect** raw CSI frames from ESP32-S3 nodes placed in a room
 2. **Extract** 8-dimensional feature vectors from sliding windows of CSI data
 3. **Contrast** -- the model learns that features from nearby time windows should produce similar embeddings, while features from different scenarios should produce different embeddings
-4. **Fine-tune** task heads using weak labels from environmental sensors (PIR motion, temperature, pressure) on the Cognitum Seed companion device
+4. **Fine-tune** task heads — *planned:* weak labels from environmental sensors (PIR motion, temperature, pressure) on the Cognitum Seed companion device. **This environmental-sensor ground-truth path is not yet implemented** (no PIR/BME280 ingestion in the training pipeline today); current task-head supervision uses the proxy/camera labels described elsewhere.
 
 ### Data provenance
 
 - **Source:** Live CSI from 2x ESP32-S3 nodes (802.11n, HT40, 114 subcarriers)
 - **Volume:** ~360,000 CSI frames (~3,600 feature vectors) per collection run
 - **Environment:** Residential room, ~4x5 meters
-- **Ground truth:** Environmental sensors on Cognitum Seed (PIR, BME280, light)
+- **Ground truth:** *Planned* — environmental sensors on the Cognitum Seed (PIR, BME280, light). Not yet wired into training; treat the PIR/BME280 references in this card as the intended design, not a current capability.
 - **Attestation:** Every collection run produces a cryptographic witness chain (`collection-witness.json`) that proves data provenance and integrity
 
 ### Witness chain
@@ -208,7 +208,7 @@ Add a second ESP32-S3 to enable cross-node signal fusion for better accuracy and
 | USB-C cables (x3) | Power + data | ~$9 |
 | **Total** | | **~$27** |
 
-The Cognitum Seed runs the ONNX models on-device, orchestrates the ESP32 nodes over USB serial, and provides environmental ground truth via its onboard PIR and BME280 sensors.
+The Cognitum Seed runs the ONNX models on-device and orchestrates the ESP32 nodes over USB serial. (Using its onboard PIR/BME280 sensors as training ground truth is planned but not yet implemented — see "Data provenance" above.)
 
 ---
 

v2/crates/wifi-densepose-train/src/dataset.rs

@@ -92,6 +92,23 @@ pub struct CsiSample {
     pub frame_id: u64,
 }
 
+impl CsiSample {
+    /// Derive the compact signal-processing feature vector for this sample
+    /// via [`crate::signal_features::extract_signal_features`] (see that
+    /// function for the layout, and [`crate::signal_features::FEATURE_LEN`]
+    /// for its length).
+    ///
+    /// Computed on demand from [`Self::amplitude`]/[`Self::phase`] — not
+    /// cached on the struct. This is the hook for folding the SOTA
+    /// signal-processing crate's amplitude/phase/PSD features (and, in a
+    /// later iteration, vitals-band power) into training; the raw vector is
+    /// returned here and is not yet fed back into the loss.
+    #[must_use]
+    pub fn signal_features(&self) -> Array1<f32> {
+        crate::signal_features::extract_signal_features(&self.amplitude, &self.phase)
+    }
+}
+
 // ---------------------------------------------------------------------------
 // CsiDataset trait
 // ---------------------------------------------------------------------------

v2/crates/wifi-densepose-train/src/lib.rs

@@ -51,6 +51,7 @@ pub mod eval;
 pub mod geometry;
 pub mod rapid_adapt;
 pub mod ruview_metrics;
+pub mod signal_features;
 pub mod subcarrier;
 pub mod virtual_aug;
 

v2/crates/wifi-densepose-train/src/signal_features.rs

@@ -0,0 +1,155 @@
+//! Hand-off layer between raw windowed CSI and the SOTA signal-processing
+//! crate ([`wifi_densepose_signal`]).
+//!
+//! Historically `wifi-densepose-signal` was listed as a dependency of this
+//! crate but never imported — the training pipeline only ever consumed the
+//! raw amplitude/phase tensors. This module wires the two together: it takes
+//! a windowed CSI observation and runs it through
+//! [`wifi_densepose_signal::features::FeatureExtractor`] to derive a compact,
+//! fixed-length feature vector (amplitude statistics, phase coherence, and a
+//! power-spectral-density summary).
+//!
+//! These derived features are the building block for a future vitals /
+//! multi-task supervision head (breathing-band and heart-rate-band power can
+//! be read off the PSD summary); for now they are produced on demand via
+//! [`extract_signal_features`] / [`crate::dataset::CsiSample::signal_features`]
+//! and are not yet fed back into the loss. Wiring them as a training target
+//! is tracked as a follow-up to the 2026-05-11 training-pipeline audit.
+
+use ndarray::{s, Array1, Array4};
+use wifi_densepose_signal::csi_processor::CsiData;
+use wifi_densepose_signal::features::FeatureExtractor;
+
+/// Length of the vector returned by [`extract_signal_features`].
+///
+/// The layout is:
+/// 1. amplitude peak
+/// 2. amplitude RMS
+/// 3. amplitude dynamic range (max − min)
+/// 4. mean of the per-subcarrier amplitude means
+/// 5. mean of the per-subcarrier amplitude variances
+/// 6. phase coherence
+/// 7. mean of the per-subcarrier phase variances
+/// 8. PSD total power
+/// 9. PSD peak power
+/// 10. PSD peak frequency (Hz)
+/// 11. PSD spectral centroid
+/// 12. PSD spectral bandwidth
+pub const FEATURE_LEN: usize = 12;
+
+/// Default centre frequency assumed when the CSI window carries no metadata.
+const DEFAULT_CENTRE_FREQ_HZ: f64 = 2.4e9;
+
+/// Default channel bandwidth (HT40) assumed when the CSI window carries no
+/// metadata.
+const DEFAULT_BANDWIDTH_HZ: f64 = 40.0e6;
+
+/// Derive a compact, fixed-length ([`FEATURE_LEN`]) signal-processing feature
+/// vector from a windowed CSI observation by running its centre frame through
+/// [`wifi_densepose_signal::features::FeatureExtractor`].
+///
+/// `amplitude` and `phase` are `[window_frames, n_tx, n_rx, n_subcarriers]`
+/// tensors (the [`crate::dataset::CsiSample`] layout). The centre frame is
+/// flattened to `[n_tx · n_rx, n_subcarriers]` (the antenna-major shape the
+/// signal crate expects) and converted to `f64`.
+///
+/// The returned values are always finite for finite input: the underlying
+/// extractors clamp degenerate cases, and any non-finite result is mapped to
+/// `0.0` so callers can rely on the vector being usable as a model feature.
+pub fn extract_signal_features(amplitude: &Array4<f32>, phase: &Array4<f32>) -> Array1<f32> {
+    let (n_t, n_tx, n_rx, n_sc) = amplitude.dim();
+    debug_assert_eq!(amplitude.dim(), phase.dim(), "amplitude/phase shape mismatch");
+    if n_t == 0 || n_tx == 0 || n_rx == 0 || n_sc == 0 {
+        return Array1::zeros(FEATURE_LEN);
+    }
+    let n_ant = n_tx * n_rx;
+    let t = n_t / 2;
+
+    let to_2d = |src: &Array4<f32>| -> Vec<f64> {
+        src.slice(s![t, .., .., ..]).iter().map(|&v| f64::from(v)).collect()
+    };
+    let amp2d = match ndarray::Array2::from_shape_vec((n_ant, n_sc), to_2d(amplitude)) {
+        Ok(a) => a,
+        Err(_) => return Array1::zeros(FEATURE_LEN),
+    };
+    let phase2d = match ndarray::Array2::from_shape_vec((n_ant, n_sc), to_2d(phase)) {
+        Ok(p) => p,
+        Err(_) => return Array1::zeros(FEATURE_LEN),
+    };
+
+    let csi = match CsiData::builder()
+        .amplitude(amp2d)
+        .phase(phase2d)
+        .frequency(DEFAULT_CENTRE_FREQ_HZ)
+        .bandwidth(DEFAULT_BANDWIDTH_HZ)
+        .build()
+    {
+        Ok(c) => c,
+        Err(_) => return Array1::zeros(FEATURE_LEN),
+    };
+
+    let feats = FeatureExtractor::default_config().extract(&csi);
+
+    let amp_mean_overall = mean_or_zero(feats.amplitude.mean.iter().copied());
+    let amp_var_overall = mean_or_zero(feats.amplitude.variance.iter().copied());
+    let phase_var_overall = mean_or_zero(feats.phase.variance.iter().copied());
+
+    let raw = [
+        feats.amplitude.peak,
+        feats.amplitude.rms,
+        feats.amplitude.dynamic_range,
+        amp_mean_overall,
+        amp_var_overall,
+        feats.phase.coherence,
+        phase_var_overall,
+        feats.psd.total_power,
+        feats.psd.peak_power,
+        feats.psd.peak_frequency,
+        feats.psd.centroid,
+        feats.psd.bandwidth,
+    ];
+    debug_assert_eq!(raw.len(), FEATURE_LEN);
+    Array1::from_iter(raw.iter().map(|&v| sanitise(v)))
+}
+
+/// Mean of an iterator of `f64`, or `0.0` if it is empty or non-finite.
+fn mean_or_zero<I: Iterator<Item = f64>>(it: I) -> f64 {
+    let (sum, n) = it.fold((0.0_f64, 0_usize), |(s, k), v| (s + v, k + 1));
+    if n == 0 {
+        0.0
+    } else {
+        sum / n as f64
+    }
+}
+
+/// Map non-finite values to `0.0` and downcast to `f32`.
+fn sanitise(v: f64) -> f32 {
+    if v.is_finite() {
+        v as f32
+    } else {
+        0.0
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use ndarray::Array4;
+
+    #[test]
+    fn zero_sized_input_yields_zero_vector() {
+        let empty = Array4::<f32>::zeros((0, 0, 0, 0));
+        let f = extract_signal_features(&empty, &empty);
+        assert_eq!(f.len(), FEATURE_LEN);
+        assert!(f.iter().all(|&v| v == 0.0));
+    }
+
+    #[test]
+    fn constant_input_is_finite_and_correct_length() {
+        let amp = Array4::<f32>::from_elem((4, 3, 3, 56), 1.5);
+        let phase = Array4::<f32>::from_elem((4, 3, 3, 56), 0.25);
+        let f = extract_signal_features(&amp, &phase);
+        assert_eq!(f.len(), FEATURE_LEN);
+        assert!(f.iter().all(|v| v.is_finite()), "features must be finite: {f:?}");
+    }
+}

v2/crates/wifi-densepose-train/tests/test_dataset.rs

@@ -458,3 +458,52 @@ fn dataloader_empty_dataset_zero_batches() {
         "iterator over empty dataset must yield 0 items"
     );
 }
+
+// ---------------------------------------------------------------------------
+// CsiSample::signal_features — the wifi-densepose-signal wiring
+// ---------------------------------------------------------------------------
+
+/// `signal_features()` must return a vector of exactly `FEATURE_LEN`, all
+/// finite, for a real (synthetic) sample.
+#[test]
+fn signal_features_have_correct_length_and_are_finite() {
+    use wifi_densepose_train::signal_features::FEATURE_LEN;
+
+    let ds = SyntheticCsiDataset::new(8, default_cfg());
+    let sample = ds.get(0).expect("sample 0 must exist");
+    let feats = sample.signal_features();
+    assert_eq!(
+        feats.len(),
+        FEATURE_LEN,
+        "signal_features() must return FEATURE_LEN ({FEATURE_LEN}) values"
+    );
+    assert!(
+        feats.iter().all(|v| v.is_finite()),
+        "all signal features must be finite, got {feats:?}"
+    );
+}
+
+/// `signal_features()` is deterministic for a given (deterministic) sample.
+#[test]
+fn signal_features_are_deterministic() {
+    let ds = SyntheticCsiDataset::new(8, default_cfg());
+    let a = ds.get(0).expect("sample 0").signal_features();
+    let b = ds.get(0).expect("sample 0").signal_features();
+    assert_eq!(
+        a, b,
+        "signal_features() must be deterministic for the same sample"
+    );
+}
+
+/// `extract_signal_features` returns the zero vector for a zero-sized window
+/// rather than panicking.
+#[test]
+fn signal_features_zero_window_is_zero_vector() {
+    use ndarray::Array4;
+    use wifi_densepose_train::signal_features::{extract_signal_features, FEATURE_LEN};
+
+    let empty = Array4::<f32>::zeros((0, 0, 0, 0));
+    let feats = extract_signal_features(&empty, &empty);
+    assert_eq!(feats.len(), FEATURE_LEN);
+    assert!(feats.iter().all(|&v| v == 0.0));
+}