Apple: same skin-temp absolute gap as #1851, but the display needs #1848 first #1853

#1852 backfills the nightly absolute on Android. Apple has the identical gap:

• skinTempC is written only by the scoring pass (IntelligenceEngine.swift:1899, skinTempC: night.nightlySkin);
• the same rolling 21-night window, so nights already outside it when the column shipped were never revisited;
• WhoopStore's DailyMetric.skinTempC (the v40 GRDB migration, twin of Room's 33→34) holds the same nulls.

The analytics half is already there and public — AnalyticsEngine.skinTempFunnel and wornNightlySkinTempC in StrandAnalytics — so this is a walker plus a fill-only write, not new maths.

Order matters, and it is not "backfill first"

Backfilling Apple now would fill a column Apple mostly does not read yet:

• the Health tile leads with the absolute (health: lead the skin-temp tile with the absolute, deviation beneath #1665) and WOULD benefit immediately;
• MetricExplorerView still sources skin temp via Repository.dailyColumn → skinTempDevC (Apple: the skin-temp explorer still leads with the deviation and ignores the Settings choice #1848), so the detail screen would not show any of it;
• the window-wide preference rule (health: apply the skin-temp preference across the window, not just the newest night #1850) and the Settings choice's reach are Android-only so far.

So #1848 first, then this — otherwise the work lands as data nothing surfaces, and the one place it does surface (the Health tile) is the place least likely to reveal a mistake in it.

What must carry across verbatim

The Android version's correctness lives in three rules, all of which were re-review findings rather than first-cut design. The Swift twin has to reproduce all three or it will be wrong in the same ways:

1. Attribute sessions by END. The night window is 54 h wide (30 h before local midnight through the next), so it contains the PREVIOUS night as well. The engine's rule is matched = allSessions.filter { tsInDay(it.end) }; handing the funnel the whole window averages two nights into one temperature and stores it as the wrong day's.

2. Page the candidates. A plain "oldest N" returns the same nights every pass, and the oldest are the ones most likely to have lost their raw samples — the sweep latches on page one and never reaches the newer nights that can fill.

3. The WHOOP 4.0 anchor comes from the CURRENT window. It is learned window-wide, and the engine's note says that is safe because the offset "cancels in the deviation" — true for a delta, not for an absolute. A re-learned anchor puts backfilled nights on a different offset from the stored ones and the chart plots two scales as one line. No anchor ⇒ decline, never the global default.

And the same non-negotiable: a single-column write that can only fill a NULL, never an upsert of a rebuilt row.