[tmm]

Follows #202. Most of the macOS issue list's remaining scroll stalls were rows paging back in: each made its SwiftUI view again and measured it (about 1.4 ms a row, 4–6 ms for a frame's first) in the frame that brought it.

What it adds

• VirtualList.maxKeptRows(_:) (a modifier, so the inits don't change) and VirtualListView.maxKeptRows: how many rows that leave the snapshot keep their views, hidden, for when they come back. The oldest are evicted first.
• A kept row that comes back is shown in its own view again, with its height and @State. If its data, its selection, or the list's width changed while it was kept, it's shown or measured again.
• Default 0: keeps none, the same as before. A kept row's own queries stay open while it's kept, so it's opt-in.
• zbugs-swift's SwiftUI issue list keeps 200 rows (about 7 MB per 100 rows).

Measured

Release builds on a 60 Hz Mac satellite, 155 open issues, the same flings-and-drags script (measure.txt, ~100 s), three alternating runs, #202 → this:

	#202	maxKeptRows(200)
frames dropped	5 / 4 / 4	6 / 2 / 5
main thread busy > 8.3 ms	238 / 259 / 302	171 / 157 / 175
busy > 16.7 ms	12 / 15 / 19	13 / 16 / 14

In Time Profiler traces, stalls ≥ 8.3 ms from rows coming into reach went from 200 spans (2.2 s) to 50 (0.6 s); all stalls ≥ 8.3 ms went from 3.5 s to 2.1 s. Row layouts a frame fell from 0.40 to 0.12, so hidden kept rows don't lay themselves out. The stalls over 16.7 ms and the dropped frames are mostly page changes (queries reopening, the window's toolbar re-rendering on each snapshot), which this doesn't change.

Tests

• keepsTheViewsOfRowsThatLeave (0, 5, 100): drives RowsLayout with row views that count what they're asked to do. Rows that page out and come back are shown and measured again only beyond the cap. 0 keeps none.
• aKeptRowComesBackAsItIsNow: data, selection, and width changed while a row was kept.
• aKeptRowComesBackWithItsState: VirtualList(...).maxKeptRows(300) in a window; rows scrolled far away and back keep their @State.

Each test fails under a deliberate break: modifier not passed through, cap ignored, stale height after a width change, stale selection. Locally on macOS: all ZeroVirtual tests (122) and Apple tests (29), the iOS Simulator builds, the macOS FrameProbeUITests, and swift format lint --strict pass.