[tmm]

Follow-up to #80 and #81: the rare paging loop #81's randomized run found, plus a regression from #80 that the same run turned up. In both, the loaded window flips between two pages without settling.

1. A short permalinked list whose row is deleted pages forever (regression in 0.6.2). A list opened at permalinkID, short enough that all its rows fit in one page and scrolled to its end, never settles after its permalink row is deleted. It alternates between r1…r14 and r0…r13 on every pass, running a page query on each one. 0.6.1 sent the list to the top and settled.

• Cause: fix(zero-swift): keep virtual lists' scroll position right when content shrinks at the bottom or their target row is deleted #80 pages forward from the first row still on screen. A forward page's start row is exclusive, so paging from a window's first row drops that row. When that window is already at the list's start, the new page loses r0, pages back from its last row, loses the end, and repeats.
• Fix: a forward page from the first row of a window that is already at the list's start is the page from the list's start (.top).

2. Scrolled past the loaded rows of short rows, the next page comes in below the viewport (the loop from #81, seed 209). When rows are shorter than estimatedRowHeight and the viewport goes past the loaded rows into estimated space, there's a blank frame and the list pages back to its start. In the worst case it flips between the two pages forever.

• Cause: fix(zero-swift): line up clamped jumps, re-pick the anchor row, and page onto the screen above a permalink #81's branch for a page that shares no row with the last one moves the space by a row estimate per index. That's right paging back, where the rows passed are the new page's and aren't measured yet. Paging forward, they're the last page's, already laid out: at 12pt rows against a 20pt estimate, the next page landed 40% further down than the rows end.
• Fix: paging forward, the space moves by the last page's laid-out height, plus an estimate for any rows between the two pages.

Tests. Both new LayoutHostTests fail on main, and each one fails without its own half of the fix:

• "deleting the row a list shorter than a page opened at, the list settles on all its rows"
• "scrolled past the loaded rows of rows shorter than estimated, the next page is on screen"

Verified locally (Linux): all 114 ZeroVirtualTests pass, and swift format lint is clean. Over 1000 seeds of the bug hunt's randomized run, with rows kept between 10 and 200pt, failures went from 10 to 8. Seed 209 and the two short-list loops are fixed, and nothing new fails.

Not fixed here: the other 8 failures. 7 are lists stuck to the bottom whose end window is short. A backward page from the window's last row leaves that row out, the space estimated after it moves the bottom, the re-pin follows it, and the list pages forward again. Upstream zero-virtual f89c468 has the same edge paging. Fixing it needs a backward page that ends at the list's end, which VirtualAnchor.backward can't represent because its startRow isn't optional. The 8th is a jump where a row moves (seed 760).

Amp-Thread-ID: https://ampcode.com/threads/T-c513a037-a257-45e0-b9a9-15c15edfe859