Sample correction range per prediction point in insulinCorrection

[ps2]

Sample correction range per prediction point in insulinCorrection

What

insulinCorrection sampled the correction range once, at the decision time, and used that single band for every predicted point. LoopKit's DoseMath samples the range at each prediction point's own time (correctionRange.quantityRange(at: prediction.startDate)), so a forecast that crosses a target-band boundary is corrected to the band actually in effect at each point.

This change matches LoopKit: look up the correction range per prediction point.

Why

With a schedule like evening 100 mg/dL → overnight 110 mg/dL, a forecast generated near the boundary extends into the overnight band, but the port corrected those points to 100 instead of 110 — systematically over-correcting into the higher band (and under-correcting into a lower one).

Replay verification against deployed Loop (real dosing history, LoopEval): this discrepancy was the dominant source of auto-bolus mismatch for a donor with an evening/overnight band change. The fix lifted exact auto-bolus agreement with the deployed app from 75.6% to 82.9% (agreement within one 0.05 U increment: 93.1% → 95.1%).

Behavior change

Only forecasts that cross a correction-range boundary are affected; with a single band (or none within the forecast horizon) the result is unchanged. The precondition message changes from "must cover date" to "must cover prediction", since the range must now cover the forecast horizon — AlgorithmInput.target timelines already do.

Testing

• New testCorrectionRangeSampledPerPredictionPoint: a flat-high forecast with the target band stepping up at t+2h requires less correction than the same forecast against the uniform low band.
• Full suite passes (142 tests).

🤖 Generated with Claude Code

https://claude.ai/code/session_01XEGCDE2D9wF3zig7uQ9PY2