Dual-timescale EWMA: recover from mid-stream level shifts (scaler, detrend)

[cboulay]

Closes #168.

Problem

Bias correction (#166) fixed cold start, but a sudden persistent level shift mid-stream (reference change, electrode impedance step, amplifier re-lock) still biases the EWMA — and downstream z-scores — for ~3·time_constant. In the scaler this reproduces the #166 collapse signature mid-session: the lagging mean biases every z-score one direction while the lagging second moment inflates the std, so features sit at a systematic z ≈ ±1 until the EWMA re-converges.

Design (as proposed in #168, decimated-detector variant)

Opt-in via EWMASettings.fast_time_constant (default None = off; detector-off output is bit-identical to before):

• A fast estimate is advanced once per chunk from the chunk mean, using the chunk's effective forgetting factor (1-α_fast)^n — no second per-sample scan, so the full-rate cost is unchanged.
• Divergence test at each chunk boundary: |fast − slow| > shift_threshold · (within-chunk residual std), with shift_hysteresis consecutive divergent chunks required (default 4·std, 2 chunks).
• Per-channel snap via xp.where: the main estimate's scan state is set so its bias-corrected value equals the fast estimate; it then continues with its own time constant. A single stepping electrode doesn't disturb the other channels.
• All detector work is elementwise on state at chunk boundaries (mean/sqrt/abs/where) — Array API portable, verified numpy/MLX parity to float32 tolerance.

Normalizing by the within-chunk residual std makes the detector robust to variance-only artifacts: a noise burst inflates the denominator, so no snap (asserted bit-identical in tests).

Scope

• EWMATransformer: detector + snap.
• AdaptiveStandardScaler: propagates the settings to both child EWMAs (mean and E[x²] snap on their respective inputs) — z-scores re-center within a few chunks of a shift.
• Detrend: inherits through EWMATransformer, nothing to change.

Tests

• Level shift: snapped estimate reaches the new level within 3 chunks; detector-off lags to end of stream.
• Per-channel: un-shifted channel bit-identical to detector-off run; shifted channel snaps.
• Stationary and variance-burst streams: bit-identical to detector-off (no false snaps).
• Scaler: signed mean z re-centers (<0.3) vs. systematic one-sided bias (>0.7) without tracking.

Defaults / tuning

shift_threshold=4.0 and shift_hysteresis=2 are conservative first choices; both only matter when fast_time_constant is set. Blend-instead-of-hard-snap and a shared (rather than per-child) detector for the scaler were considered and deferred — noted as follow-ups in #168 if needed.

[cboulay]

I don't plan to merge this anytime soon. It's a solution for a problem that I haven't really encountered yet and thus likely adds unnecessary complication.