feat: cpu-first real-csi inference vertical slice

[beastbroak30]

Problem

RuView had no CPU inference path for real public CSI datasets, no dataset
preparation pipeline, and no honest profiling harness. This patch adds the
first vertical slice: dataset validation, preprocessing, CPU profiles, a
benchmark binary, and documentation — all without downloading data or training
weights.

Datasets

Dataset	Source	License	Status
MM-Fi (primary)	ybhbingo/MMFi_dataset	CC-BY-NC-4.0	Manifest + loader ready; download manually
Wi-Pose (secondary)	NjtechCVLab/Wi-PoseDataset	Research use	Manifest template only; .mat → .npy conversion documented but not wired

Both datasets require manual download. Files are never committed to Git.

Implementation

New modules:

• public_dataset.rs — DatasetManifest, PreparedSplit, ProcessedSampleMeta, SplitPolicy
• real_csi.rs — RealCsiPipeline (6-stage: shape validation → NaN rejection → phase sanitization → Hampel denoising → per-pair standardization → subcarrier conversion)
• profile.rs — CpuProfile (5 variants, 2 implemented), ModelMetadata with enforced honesty separation

New binaries:

• prepare_public_dataset — validates .npy headers, measures native width, writes dataset_manifest.json + split.json
• cpu_inference_bench — machine-headered P50/P95/throughput reports, --compare FP32 vs INT8

Modified:

• inference.rs — InferenceOptions gains profile, precision, deterministic
• onnx.rs — session builder honors intra-op thread count
• dataset.rs — MmFiDataset::discover_subset for subject-filtered discovery
• wifi-densepose-signal — re-exports hampel_filter and HampelConfig

Methodology

1. Preprocessing (real-csi-v1): fail-closed (no NaN/inf repair), documented interpolation math for 114→56 conversion, per-sample provenance.
2. Splits: subject-disjoint by default (Fisher–Yates, seeded, self-checked). Environment/recording/cross-dataset splits documented as future requirements.
3. Benchmarks: criterion bench for preprocessing; ONNX inference bench with warmup, multiple iterations, JSON reports.
4. Honesty: ModelMetadata::validate() refuses to claim accuracy without real validation. All cells in the results table are "not run (no dataset locally)".

Results

No accuracy numbers claimed. All infrastructure verified:

• Tests: 355 new (train 283, nn 97, signal 496)
• Clippy: zero warnings in new code
• Smoke: both bins ran successfully on Windows x86-64

Limitations

1. No MM-Fi/Wi-Pose files downloaded here — no accuracy numbers exist.
2. Public-dataset accuracy ≠ RuView ESP32 accuracy (different chipsets, antennas, rooms).
3. 114→56 interpolation loses frequency resolution.
4. MM-Fi is CC-BY-NC-4.0: no commercial deployment of MM-Fi-trained weights.
5. x86 int8 numbers do not transfer to edge-SoC int8 latency.

Reproducing

# tests
cargo test -p wifi-densepose-train --no-default-features --lib --tests --bins
cargo test -p wifi-densepose-nn --lib --tests --bins
cargo test -p wifi-densepose-signal --no-default-features

# smoke bins
cargo run -p wifi-densepose-nn --bin cpu-inference-bench -- --profile cpu-micro-fp32 --threads 1