[ryanbr]

Closes #1448.

The defect

The Deep Timeline's windowed rMSSD took successive differences across beats the artifact filter had
removed. Dropping a beat joins two intervals that were never adjacent, and their difference is a
splice, not a physiological delta — the spurious large delta this analyzer's gap-aware pair (#204/#195)
exists to exclude.

The nightly analyze() already used cleanRRGapAware + rmssdGapAware. rollingRmssd used
cleanRR + rmssdRaw three lines away in the same file — on both platforms.

	Swift	Kotlin
nightly analyze	gap-aware ✅	gap-aware ✅
rollingRmssd	raw ❌	raw ❌

The twins agreed with each other while both were wrong, which is exactly why no parity check ever
surfaced it. It bites hardest on the messiest data — on a WHOOP 4.0 that is most of it (#1008) — and the
trace exists specifically to be an honest windowed rMSSD (#803) rather than the raw-interval chart it
replaced.

Why this stays narrow

• CleanSeries.nn is documented as byte-identical to cleanRR over the same input, so the
  minBeatsPerWindow survivor gate does not move.
• rmssdGapAware is identical to rmssdRaw when there are no gaps, so a clean window is
  unchanged. Only windows that actually lost a beat move.

The one visible consequence: a window whose survivors share no adjacent pair now emits nothing
rather than a number assembled entirely from splices. That is honest absence, and consistent with the
existing sparse-window gate that already prefers no point over a noisy one.

Tests

Mirrored fixtures per platform, built on the range filter so the drop is deterministic rather than
depending on Malik's neighbour median:

• five 1000 ms beats, an out-of-range 2400 ms beat, five 1150 ms beats → the 150 ms step is pure splice,
  and the emitted value goes from a filter-invented 50.0 → 0.0
• the control is the important half: the same two runs with no beat between them still read exactly
  50.0, proving this is not a numbers-move for clean data

Verification

• Android 4,070 tests / 0 failures (--rerun-tasks, both new vectors confirmed by name)
• i18n and doc-comment gates clean
• Both changed files sit under Packages/ and Android, so swift-packages executes the Swift twin
  — no app-target Swift touched, so no app-build gate needed
• Not device-tested: this is pure analytics with no BLE path, and the fixtures cover it