Sport Mode, part 3/5: G7SensorKit — watchOS target + radio acquisition fixes

[ps2]

What this is

The G7SensorKit slice of Jeremy Barnum's Sport Mode line (standalone watch). Part 3 of a
bottom-up landing: LoopKit (#612) → RileyLinkKit (#18) / G7SensorKit → OmnipodKit → Loop →
workspace pin bump.

Draft: design review wanted before this goes near next-dev.

Eleven commits, rebased onto current next-dev (fb87aa9). One real conflict, resolved — see
below.

The changes, in two groups

1. watchOS support (4 commits) — additive

• New G7SensorKit-watchOS framework target and its shared scheme.
• watchOS source guards and a transport seam.
• Deployment target 10.6, to match LoopKit.
• Pins the iOS target's SDKROOT. Without it, an implicit-dependency hijack lets the watchOS
  target satisfy an iOS dependency — worth a look, since it touches the existing iOS target.

2. G7 radio acquisition work (7 commits) — behavioral, iOS included

This is the half that deserves scrutiny, because it changes stock iOS G7 behavior:

• "Fix the 20–40 minute acquisition outage: arm the scan whenever not connected." A real
  acquisition bug with a real fix, and the most valuable commit here for non-watch users.
• G7 radio census — makes the acquisition mechanism observable; exposes sensor sightings as a
  signal rather than prose, and adds [domain#code] on connect failures.
• Watchdog delivery baseline seeding, so the first check isn't infinity.
• "G7 acquisition doorways settled — lab toggles hardcoded true." These were lab toggles that
  are now hardcoded on. Whether that is the right shape for upstream, or whether they want to stay
  configurable, is a design call for this review.

Conflict resolved (please check this one)

G7BluetoothManager.init conflicted. next-dev (#63) had extracted central-manager creation into
a makeCentralManager(queue:) factory seam so tests can substitute a manager without the state
restoration option; Jeremy's commit had inlined #if os(iOS) guards at the same call site.

I kept your factory and moved the platform distinction inside it:

func makeCentralManager(queue: DispatchQueue) -> CBCentralManager {
#if os(iOS)
    return CBCentralManager(delegate: self, queue: queue, options: [CBCentralManagerOptionRestoreIdentifierKey: "com.loudnate.CGMBLEKit"])
#else
    // watchOS has no CoreBluetooth state restoration; the watch host owns reconnect policy.
    return CBCentralManager(delegate: self, queue: queue, options: nil)
#endif
}

This preserves both intents and keeps G7CGMManagerTests.TestBluetoothManager's override of
makeCentralManager working. It is the one place in this PR where the resulting code matches
neither side verbatim, so it is the thing to check hardest.

Everything from #63 — the suspected-session-end grace period, its persistence across termination,
and its tests — is carried forward untouched.

Verification

• xcodebuild -workspace LoopWorkspace.xcworkspace -scheme LoopWorkspace against stock
  next-dev (stock LoopKit, this branch + the RileyLinkKit branch swapped in): BUILD
  SUCCEEDED.
• -scheme G7SensorKit-watchOS -destination 'generic/platform=watchOS Simulator': BUILD
  SUCCEEDED.
• Content diff against Jeremy's G7SensorKit tip is exactly the intervening next-dev commit plus
  the conflict resolution above.

The G7SensorKit test suite has not been run yet.