BFI research tools 0.1.0-rc.1

This source prerelease adds the BFI capture and offline CSI/BFI research tools from PR #2055.

Included

• Strict BFI decoding, independent tshark comparison, capture/traffic coverage analysis and bounded Linux capture runners.
• Numeric data preparation, bounded LSTM/GRU training, static and temporal controls, frozen checkpoint evaluation and Codex review jobs.
• A pinned public BFI dataset manifest/downloader and a separate VHT MU 3x1 report-body decoder.
• Proposed ADRs 364–367, Linux CPU regression CI, and mt76 patch attribution with the BSD-3-Clause-Clear license.
• Wasmtime 36.0.16 and matching compiler-component patches address RUSTSEC-2026-0316.

Measured evidence and limits

• All 162 retained reports from two BE550/iPhone/MT7927 captures matched the independent angle/control/SNR oracle. Traffic-period gaps of 8.392 and 6.437 seconds still fail the proposed continuity target.
• The public sample decoder matched 13,104 angle integers across 14 bodies. Whole-packet admission and public BFI activity training remain incomplete.
• Sequence models and static controls both reached 56/56 recordings on the custom CSI validation split. This does not demonstrate gait learning, KIT-style BFI identity recognition or state-of-the-art performance. The holdout remains unscored at the recorded research checkpoint.

All 64 capture tests and 141 learning tests passed in Linux CPU CI, including the symlink case skipped on Windows. Local deterministic synthetic proof, documentation, integrity and independent publication reviews also passed.

See capture usage, learning usage and measured results.

Raw captures, subject datasets, weights, credentials and private logs are excluded. This release distributes research source; it does not promote a production model or publish a package.