feat(sleep): promote Sleep V2 to the default staging engine + tune deep boundary

[tanarchytan]

Promotes the transparent Sleep V2 recipe (from #600) from opt-in to the default staging engine, and tunes its deep boundary. V1 stays available behind the flag. This is the staging half of the sleep-tracking work; the companion detection fix (recovering nights that were dropped whole on an HR spike) already merged in #268, so both V1 and V2 now detect every night.

Before / after

Per-epoch, cross-subject leave-one-subject-out, run through the real shipped stager (not a reimplementation) on two public wrist datasets with PSG ground truth: AAUWSS (13 subjects, E4 wrist + gold PSG-ECG R-R + AASM labels) and Walch 2019 sleep-accel (31 subjects). BEFORE is the current default (V1 staging). AFTER is what this PR makes the default (V2 staging).

metric	BEFORE (V1 default)	AFTER (V2 default)
Cohen kappa, AAUWSS	0.039	0.415
Cohen kappa, sleep-accel	0.089	0.352
sleep/wake, AAUWSS	86.4 %	88.7 %
sleep/wake, sleep-accel	88.8 %	91.3 %
4-class acc, AAUWSS	44.3 %	59.4 %
4-class acc, sleep-accel	50.6 %	58.1 %
deep recall, AAUWSS	1 %	62 %
deep recall, sleep-accel	9 %	60 %

V1's near-zero kappa and 1 % deep recall are the flat, deep-poor hypnograms users see today. V2 recovers deep and REM. V1 stays available behind the same flag for anyone who prefers it.

V2 shipped opt-in because it had only n=1 validation. The 44-subject benchmark above is the multi-subject validation it lacked, and it shows V2 strictly dominates V1 on every metric on both datasets.

The deep-boundary tune

On the same benchmark, V2's deep boundary is tuned with parity-safe constant edits (set a-priori from physiology and population base rates, none fit to labels):

• deepGateThresh 0.20 to 0.25 (recover the deep recall the tighter edits shed)
• deep emission -1.4*zhvv - 0.2*zhrv - 0.3*zmvv to -1.1*zhvv - 0.5*zmvv (drop the mean-HR term; HR-variability and movement carry deep)
• deep transition row deep .90 to .86 (renormalises to rem .007 / light .126 / awake .007)

Isolated impact (untuned to tuned V2, same detection): kappa +0.018 (AAUWSS) / +0.027 (sleep-accel), 4-class +1.9 / +2.6, sleep/wake flat, deep precision up. Nothing slid.

Parity and tests

Byte-identical Kotlin and Swift: SleepStagerV2.kt / SleepStagerV2.swift match on deepGateThresh, the transition row, and the deep emission. The promote resolves a missing preference to the V2 default on both platforms (Android getBoolean(KEY, true); macOS an explicit nil-key check before bool(forKey:)), the two @AppStorage defaults are true, and the Settings and Test Centre copy is updated so it no longer describes V2 as opt-in / off by default. The V2 tests are property-based (segments tile the window, canonical stage labels, deep and REM present, cycle-prior monotonicity) and pass unchanged. Kotlin compiles and tests pass; the Swift package build needs macOS, so please confirm the swift test leg in CI.

Scope and notes

• Pure analysis switch. Sleep DETECTION, scoring, and the V1 code path are untouched; the same detected window is simply staged by V2. Model-agnostic (WHOOP 4 and 5).
• The per-epoch V2 coefficients remain a fixed, transparent recipe. This is not a trained model.
• With iOS: write sleep stages, 1-min heart rate, and workouts back to Apple Health #249's Apple-Health sleep-stage writeback on iOS, making V2 the default means the richer V2 deep/REM stages now flow to Apple Health.