sleep-edf: declare and apply microvolt input scaling (v0.2)

[bondingelectron]

Summary

• scale epochs to microvolts in load_subject_recording (×1e6, float32), matching the academic convention for EEG (MOABB / braindecode expose µV-scale arrays) and the bciciv-2a v0.2 precedent — keeps input units consistent across the bench
• declare "input_unit": "microvolts" in the prompt's frozen preprocessing dict and note µV in the input shape contract
• add 4 preprocessing contract unit tests (label mapping incl. stage 3+4 → N3 merge, V→µV scaling, interface constants, training hyperparameters) and wire them into the required CI matrix
• document the unit in both READMEs and bump the breaking task version from 0.1 to 0.2

Background

The scorer returned MNE's SI-unit (volt) arrays while the prompt's self-contained preprocessing dict said nothing about units, and the shape contract exemplified inputs with torch.randn (unit scale). This is not a correctness bug — the scorer trains and tests on the same pipeline, so scores stay fair and reproducible — but amplitude-sensitive architectures (fixed-eps log-power, unnormalized saturating nonlinearities) behave differently at 1e-5 scale, and an agent inspecting the raw EDFs with non-MNE tooling sees µV (the EDF physical unit), which contradicts what the scorer actually feeds the model.

Validation

• 4 new contract tests pass; bciciv-2a contract tests still pass (5/5)
• Real-data smoke on pinned subject-0 recording: (841, 1, 3000) float32, label distribution 188/58/250/220/125 (W/N1/N2/N3/REM) unchanged, std = 22.88 µV (was 2.29e-5 V), abs max 192 µV
• bp-bench validate sleep-edf, docs smoke, and the 208-test Node suite all pass

Score compatibility

This is a breaking scorer change. Scores produced with sleep-edf v0.1 should not be mixed with v0.2 scores. For normalization-based architectures the difference is expected to be within run-to-run noise; amplitude-sensitive architectures benefit materially.

[GTC2333]

Verified the V→µV conversion on the pinned Sleep-EDF data: epoch shape and labels remain unchanged, while amplitudes scale exactly as intended. The four new preprocessing-contract tests and all required CI checks pass. Approved for merge. Non-blocking follow-up: please align the task README header from Version 0.1 to 0.2 and explicitly note that v0.1/v0.2 scores are not comparable.