[ryanbr]

Addresses #1702, on the decision to match Android's window.

What was actually wrong

The two platforms rendered the same section from different data. Android windows to today − 13 at start of day and says "14 days" in the header. iOS drew from repo.workoutRows(), whose default is days: 4000 — effectively all-time — under a heading that says "Latest", trailed by a count of every workout ever recorded.

One account read two ways: someone who last trained three months ago saw no section on Android and six stale sessions on iOS.

Reading the call sites suggests iOS's behaviour was never a product decision. Android keeps two sets — recentUnion for the feed, separate all-time queries for the Data Sources counts. iOS has a single workouts array serving both, and the section simply used what was already there.

Why the window is applied at the section

This is the part worth reviewing carefully. workouts is shared:

• the Data Sources "Apple Health · N workouts" count (TodayView.swift:3939)
• the HR chart's sport glyphs (workoutSpans)

Both are all-time by design. Windowing the array at its source compiles, looks correct, and quietly shrinks two unrelated numbers on the same screen. So TodayView.recentWorkoutsFeed windows at the section and the shared array is left alone.

Also

• Trailing header now describes the window on both. "14 days" already existed in the catalogue in all ten locales, so there is no new copy and no xcstrings work.
• "\(workouts.count) total" is gone — it counted all-time while showing six.
• Tile caps stay 4 (Android) / 6 (iOS), deliberately. That difference is layout, not content: Android is single-column full-width on purpose (the 2×2 grid truncated durations on narrow phones, Future-dated strap (#324): surface a banner + bad-clock diagnostics #332), iOS uses an adaptive grid. Same vertical budget; unifying it would change one platform's density for no reader benefit.

Tests

TodayWorkoutsWindowTests (Swift, runs under app-build) pins the boundary, not the happy path: a session exactly on the cutoff is kept, one second earlier is dropped, and filtering must not reorder. An off-by-one there silently drops a session the other platform still shows — the same class of drift that produced the issue.

TodayWorkoutTapTest gains a parity assertion that runs in ordinary CI and explicitly checks the Data Sources count still reads the unwindowed array. The trap above is guarded, not merely sidestepped once.

Verified

4,713 Android tests, 0 failures with --rerun-tasks --no-build-cache; Tools/i18n_audit.py --ci clean. The Swift half compiles only under app-build — dispatching it, and I will post the result here before this merges.

Note for review

This changes iOS behaviour: the section now disappears when you have not trained in two weeks, matching Android. That is the consequence of the chosen option, not a side effect.