Focus24 Kernel

After parity is proven, each lands as a separate tested change (focus24/docs/paper/focus24-kernel.pdf §6):

1. PERCLOS is clamped to 0–1 while the eye metrics report 0–100, so anything above 1% saturates. The Rust unit types already prevent this; the fix is to set the thresholds accordingly.
2. The circadian tracker is never instantiated, so its score stays empty.
3. rPPG samples at 6.7 Hz into a 300-sample buffer, giving a 45 s window and a filter edge above Nyquist.
4. The neck model advances one second per call at 4 Hz, so fatigue accumulates four times too fast.
5. The state machine applies the exit bound while inside a state, so the top band oscillates between 85 and 88.
6. There is no gaze calibration, so an off-axis monitor counts as looking away.

Items 1 and 3 affect the reported numbers most.