[ps2]

LoopKit depends on LoopAlgorithm today only for LoopQuantity and LoopUnit, so an app that wants LoopKit without the algorithm (Trio) still has to build LoopAlgorithm. This changes LoopAlgorithm's public API to plain Doubles in fixed units. LoopKit can then own its unit types, and neither package depends on the other, with no new shared package.

This is the first of a set of changes. LoopKit and Loop PRs follow, and Loop becomes the only place that converts between LoopKit's types and LoopAlgorithm's (see the plan below).

API changes

	Before	After
Glucose	quantity: LoopQuantity	mgdl: Double (GlucoseValue, GlucoseEffect, PredictedGlucoseValue, GlucoseChange)
Carbs	quantity	grams: Double (CarbEntry, CarbStatus, AbsorbedCarbValue.observedGrams etc.)
Velocity	quantity	mgdlPerSecond: Double
Sensitivity	sensitivity: [AbsoluteScheduleValue<LoopQuantity>]	sensitivityMgdlPerUnit: [AbsoluteScheduleValue<Double>]
Target	target: [AbsoluteScheduleValue<ClosedRange<LoopQuantity>>]	targetMgdl: GlucoseRangeTimeline (ClosedRange<Double>)
Suspend threshold	suspendThreshold: LoopQuantity?	suspendThresholdMgdl: Double?
Momentum limit	momentumVelocityMaximum: LoopQuantity?	momentumVelocityMaximumMgdlPerMinute: Double?

The plain-number inputs carry their unit in the name, both on AlgorithmInput and in the public overloads (generatePrediction, PrecomputedInsulinInput.withEffects/build, the glucoseEffects and carb-effect overloads). That way an mmol/L value passed by mistake stands out at the call site instead of compiling unnoticed. The Loop PR adds an end-to-end mmol/L test as well.

• LoopQuantity and LoopUnit are now internal. Internally the math still uses them, so results are unchanged.
• The LoopAlgorithm struct is renamed Algorithm, so LoopAlgorithm.X always names the module. Call sites become Algorithm.run(input:) and Algorithm.generatePrediction(...).
• GlucoseSampleValue drops trendRate and condition; the algorithm never read them. FixtureGlucoseSample.condition and the internal GlucoseCondition go too.
• run converts the input schedules to quantities once at the start.
• Helpers apps don't call (retrospective correction types, insulinCorrection, recommendTempBasal, etc.) are internal. The fixture JSON format is unchanged.
• LoopKit's tests of LoopMath (prediction, momentum blending, ICE − carb effect, decay), date rounding, dose annotation and IOB move here with their fixtures, since they test this package's code. The dose tests read LoopKit's pump event dose format through a small test-only loader, with LoopKit's Walsh curve and daily basal schedule expansion ported for the tests.

Plan

LoopWorkspace next-dev pins LoopKit/LoopAlgorithm e3f5e3c, so merging this to main doesn't affect it. The workspace bump that adopts this change needs these PRs too:

• Own the shared value types instead of depending on LoopAlgorithm LoopKit#674: owns LoopQuantity, LoopUnit and the shared value types under the same names. It drops its LoopAlgorithm dependency and builds and tests on its own.
• Use LoopKit's value types and bridge to LoopAlgorithm Loop#2635: the bridge between LoopKit's types and the algorithm, plus an end-to-end mmol/L test.
• Plugins that use these types through import LoopAlgorithm. LoopKit keeps the names, so each PR just switches to import LoopKit:
  • Import LoopKit instead of LoopAlgorithm CGMBLEKit#230, Import LoopKit instead of LoopAlgorithm G7SensorKit#106, Import LoopKit instead of LoopAlgorithm MinimedKit#43
  • Import LoopKit instead of LoopAlgorithm NightscoutService#45, Import LoopKit instead of LoopAlgorithm NightscoutRemoteCGM#41
  • Import LoopKit instead of LoopAlgorithm LibreLoop#32, Import LoopKit instead of LoopAlgorithm LibreTransmitter#60
  • Import LoopKit instead of LoopAlgorithm EversenseKit#10, Import LoopKit instead of LoopAlgorithm MedtrumKit#8
  • Import LoopKit instead of LoopAlgorithm dexcom-share-client-swift#60 (ShareClient), Import LoopKit instead of LoopAlgorithm LoopOnboarding#40
  • Import LoopKit instead of LoopAlgorithm loopandlearn/OmnipodKit#206, which Trio shares
  • Import LoopKit instead of LoopAlgorithm TidepoolService#47, which also adds a file-private copy of a date helper it used from LoopAlgorithm
• LoopKit/LoopWorkspace: the bump, once the above merge.

Testing

• swift test: 214 tests pass (main's tests plus the 26 moved from LoopKit).
• With the companion LoopKit and Loop changes, I ran closed loop in the simulator. I saved the input of 17 algorithm runs (automatic boluses, temp basals, manual boluses with and without notices). Replaying each through main as of e3f5e3c (before Guard + Runway, second version; slow retrospective correction experiment (off by default) #16) and through this change gave identical forecasts, effects, IOB/COB and recommendations (within 1e-9). After rebasing onto Guard + Runway, second version; slow retrospective correction experiment (off by default) #16, its experiment tests pass unchanged apart from the unit updates.