Automatically filter out duplicate-sync sleep data

[gsbernstein]

What

Some sources (notably Oura) have a "double sync" bug: syncing a night, then falling back asleep and syncing again, re-writes the same stretch of stage samples a second time instead of only writing the newly-recorded tail. Left alone, that inflates every duration total the app computes (Sleep Bank balance, per-night totals, source comparisons) by roughly 2x for the affected night.

History

This PR originally added an in-app UI to detect these duplicates and delete them from HealthKit on confirmation (a histogram + timeline + draggable divider + delete button). That turned out to be fundamentally impossible: HKHealthStore.deleteObjects(of:predicate:) (and delete(_:)) are documented to only ever delete objects the calling app itself wrote — a third-party app can never delete another app's HealthKit samples, no matter what write authorization it has. Since real duplicates always come from the other source (Oura), not Bedger, that delete call matched zero of Bedger's own objects and silently "succeeded" with nothing actually removed.

That delete-based feature (detection UI, histogram/timeline views, and the write-authorization/delete plumbing in HealthKitManager) has been reverted. In its place, SleepSampleDeduplicator filters the older, superseded sync out of Bedger's own calculations instead of trying to touch HealthKit at all.

An earlier version of this picked which sync to drop by comparing total covered duration of overlapping "generations" — but that biases toward whichever sync happens to produce longer/coarser stage segments, not necessarily the newer one. It's now based on "Date Added" instead — the same signal a person would use looking at a sample in the Health app.

What this PR does now

• HealthKitCreationDateReader reads a sample's internal creationTimestamp via undocumented Key-Value Coding (guarded by a small Objective-C safe-KVC trampoline, KVCSafeAccessor, wired in via a bridging header) — the only way to know when HealthKit received a sample, since startDate/endDate only describe when the sleep happened. This is the same reader used (and already debugged) in the original delete-feature attempt, now used purely for internal filtering — never surfaced in any UI or write path, so if it ever stops working on some future OS, filtering just silently no-ops for that sample instead of anything user-visible breaking. No HealthKit write authorization needed; it only reads a property on samples already covered by existing read access.
• SleepSampleDeduplicator.deduplicate(_:) groups raw HKCategorySamples by (night, source). For a group that shows the structural signature of a duplicate re-sync — same-source samples overlapping in time, which never happens within a single healthy sync since stage segments are sequential — it finds the largest gap between distinct creation timestamps across the whole group (not just the contested overlap span) and drops every sample created before that cutoff. Samples with no readable creation date, or groups without enough distinct timestamps to place a meaningful cutoff, are always kept — never guessed away.
• Wired into HealthKitManager.processSleepSamples, before samples become SleepSessions, so every downstream consumer (Sleep Bank, per-night totals, source comparison view, etc.) benefits automatically. Preserves the original reverse-chronological sample ordering that SleepDayGroup/LastNightCard rely on for wake/bed times.

Notes for verification (macOS/Xcode only — see AGENTS.md)

• The bridging header (Bedtime/Bedtime-Bridging-Header.h) is wired into both Debug and Release build settings (SWIFT_OBJC_BRIDGING_HEADER) in project.pbxproj; the .h/.m/bridging-header files live under the existing file-system-synchronized Bedtime/ group so Xcode should pick them up automatically without a manual "Add Files" step.
• No new authorization prompts — this only reads a KVC property on already-read data.
• To exercise this, seed HealthKit with two overlapping batches of stage samples for the same source/night at different times (e.g. write a night via DebugDataGenerator, then write an overlapping/extended second batch a bit later under the same source) and confirm the Sleep Bank total and SleepDayGroup duration reflect only the newer batch.