fix(train): wire wifi-densepose-signal into the pipeline + correct MODEL_CARD env-sensor claim (audit #1/#2/#3)

[ruvnet]

Summary

Three of the seven findings from the 2026-05-11 training-pipeline audit (Stuart's Stage-A audit; tracked in PR #535):

#	Finding	This PR
1	Vitals/CSI signal features absent from the training pipeline	✅ new signal_features module that extracts them
2	wifi-densepose-signal is a phantom dep of wifi-densepose-train (listed, never imported)	✅ it is now an actually-used dependency
3	PIR/BME280 env-sensor ground truth in MODEL_CARD.md has no implementation	✅ marked planned / not-implemented
5	README "92.9% PCK@20" overclaim	(fixed in #535)
4, 6, 7	proof.rs uses synthetic data only · 56-sc/1-NIC default · multi-band 168-sc not in config	not in this PR — see "Follow-up"

#1 + #2 — wifi_densepose_train::signal_features

wifi-densepose-signal (the SOTA signal-processing crate) was in wifi-densepose-train/Cargo.toml but never used — and the training samples carried only raw amplitude/phase, no derived signal features. New module:

• extract_signal_features(amplitude: &Array4<f32>, phase: &Array4<f32>) -> Array1<f32> — takes a windowed CSI observation ([T, n_tx, n_rx, n_sc]), flattens the centre frame to the antenna-major [n_tx·n_rx, n_sc] shape the signal crate expects, builds a wifi_densepose_signal::csi_processor::CsiData, and runs wifi_densepose_signal::features::FeatureExtractor over it.
• Returns a fixed-length FEATURE_LEN = 12 vector: amplitude peak/RMS/dynamic-range + mean-of-per-subcarrier amplitude mean/variance, phase coherence + mean phase variance, and a PSD summary (total power, peak power, peak frequency, centroid, bandwidth). Non-finite results are mapped to 0.0; degenerate (zero-sized) windows return the zero vector instead of panicking.
• Convenience method CsiSample::signal_features(&self) -> Array1<f32>.
• Scope note: the vector is produced on demand and is not yet fed back into the loss. It's the hook for a future vitals / multi-task supervision head — breathing-band and heart-rate-band power can be read off the PSD summary. Wiring it as a training target is the documented follow-up. But wifi-densepose-signal is no longer a ghost dep, and the signal features the audit said were "absent" are now computed and accessible on every sample.

#3 — docs/huggingface/MODEL_CARD.md

The card presented PIR/BME280 weak-label fine-tuning on the Cognitum Seed as a current capability. There is no env-sensor ingestion in the training pipeline. Marked it as planned/not-implemented in the training-steps list, the data-provenance section, and the hardware section.

Tests

• cargo check -p wifi-densepose-train --no-default-features — clean (only pre-existing wifi-densepose-signal dead-code warnings).
• cargo test -p wifi-densepose-train --no-default-features — all existing tests green, plus 5 new tests pass:
  • signal_features::tests::{zero_sized_input_yields_zero_vector, constant_input_is_finite_and_correct_length}
  • tests/test_dataset.rs::{signal_features_have_correct_length_and_are_finite, signal_features_are_deterministic, signal_features_zero_window_is_zero_vector}
• No change to tch-gated modules, no change to the model input shape or the deterministic proof — purely additive.

Follow-up (the remaining audit findings)

• Fixed indentation issue on diagram in README.md #4 — proof.rs uses SyntheticCsiDataset: that's correct for a deterministic proof (a reproducible source is the point); the real gap is no smoke-test of the real .npy MmFiDataset loader path. Recommend adding a round-trip test (generate synthetic → write .npy → load via MmFiDataset → assert shapes) — left out here to keep this PR focused.
• Error #6 / unable to access wifi-densepose REST API on macos #7 — config defaults at 56-sc/1-NIC and no multi-band/168-sc preset: recommend a TrainingConfig::for_mesh(native_subcarriers, n_nodes) constructor + docs. Separate PR.
• Vitals supervision head — once signal_features is consumed: add an optional vitals-target field + a small head/loss term. The training-guide.md (delivered to the pulse-seed handoff) describes the end-to-end path.

🤖 Generated with claude-flow