feat(train): TrainingConfig subcarrier-layout presets + real MmFiDataset loader test (audit #4/#6/#7)

[ruvnet]

Summary

Closes the remaining doable findings from the 2026-05-11 training-pipeline audit (after #535 fixed #5 and #536 fixed #1-extract/#2/#3):

#	Finding	This PR
6	CSI format default = 56-sc / 1 NIC — non-MM-Fi shapes require hand-overriding two fields	✅ TrainingConfig::for_subcarriers + named presets + docs
7	Multi-band 168-sc mesh (ADR-078) not reflected in the training config	✅ TrainingConfig::multiband_168() preset + multi-NIC mapping documented
4	proof.rs uses SyntheticCsiDataset not real data	✅ reframed + tests/test_real_loader.rs covers the real on-disk path

#6 / #7 — TrainingConfig subcarrier-layout presets

num_subcarriers: 56 / native_subcarriers: 114 were the only documented shape; using any other CSI source meant overriding both fields by hand. New, in config.rs:

• TrainingConfig::for_subcarriers(native, target) — explicit native→model resample.
• TrainingConfig::mmfi() (114→56, == default()), ht40_192() (≈192-sc ESP32 HT40 → 56), multiband_168() (168-sc ADR-078 multi-band mesh → 56).
• Field docs now list the common native counts (MM-Fi 114, HT20 56, HT40 ≈192/114, multi-band mesh 168) and point at the presets, plus a note that a 2–3-node CSI mesh currently maps onto [T, n_tx, n_rx, n_sc] via extra n_rx until a dedicated node dimension lands.
• The model's input width stays num_subcarriers — the presets only change the resampling input, not the architecture.

#4 — real on-disk loader test

A deterministic proof must use a reproducible source, so verify-training correctly stays on SyntheticCsiDataset — the real gap was that nothing exercised MmFiDataset's .npy disk-loading path. New tests/test_real_loader.rs:

• writes synthetic CSI (deterministic, no RNG) to .npy files in the MmFiDataset::discover layout (<root>/S01/A01/{wifi_csi.npy, wifi_csi_phase.npy, gt_keypoints.npy}),
• loads it back via discover + get, asserting CsiSample shapes ([T,n_tx,n_rx,n_sc], keypoints [17,2], visibility [17]) and finiteness,
• covers the subcarrier-interpolation branch (target 28 < native 56 → resampled output shape) and the empty-root case.
• Adds ndarray / ndarray-npy as [dev-dependencies] for fixture writing.

Tests

cargo check -p wifi-densepose-train --no-default-features — clean (only pre-existing ruvector-* / wifi-densepose-nn warnings). cargo test -p wifi-densepose-train --no-default-features — all existing tests green, plus the 3 new test_real_loader.rs tests and the updated config.rs doctest (now exercises ht40_192() / for_subcarriers) pass. Purely additive — no model-shape change, no tch-module change, deterministic proof untouched. (An unrelated 109-line Cargo.lock drift a cargo check introduced was reverted, so the diff is clean.)

Audit scoreboard

#1 (extract) ✅ #2 ✅ #3 ✅ (#536) · #4 ✅ #6 ✅ #7 ✅ (this PR) · #5 ✅ (#535). The one not closed: #1's training-target half — wiring the extracted signal_features into a vitals/multi-task supervision head — is blocked on a vitals-labeled dataset (MM-Fi/Wi-Pose carry no vitals labels; an ESP32 paired capture would). The end-to-end path is in the training-guide.md delivered to the pulse-seed handoff. So all 7 findings are addressed/reframed; one has a remaining follow-up gated on data.

🤖 Generated with claude-flow