[exoego]

Problem

With the new engine (experimentalPagination or a page layout), a node with minPresenceAhead can be left at the bottom of a page with only a line or two after it. This happens when the next sibling is a container that stacks its children vertically, like a table made of View rows. The engine breaks the page inside that container.

The legacy engine moves the node to the next page instead. #3504 promised the same thing for the new engine. No page break may fall inside the window after the node. If the window does not fit, the node moves to the next page.

Repro

<Document>
  <Page size="A6" style={{ padding: 20 }} experimentalPagination>
    <View style={{ height: 330 }} />
    <Text minPresenceAhead={60}>Heading</Text>
    <View>
      {Array.from({ length: 20 }, (_, i) => (
        <Text key={i}>{`row ${i + 1}`}</Text>
      ))}
    </View>
  </Page>
</Document>

Case	Page 1	Page 2
Legacy engine	(no heading)	Heading, row 1, …
New engine	Heading, row 1	row 2, …
New engine, rows not wrapped in a View	(no heading)	Heading, row 1, …
New engine, with this fix	(no heading)	Heading, row 1, …

Cause

#3504 checks the window at three places. These are break candidates, leaf splits, and row slices. A split inside a column is not checked.

fit/row.ts declines a row slice inside a window. fit/column.ts has no such check. The column's children are filled by a separate fill that does not know about the outer window. So the outer fill accepts the split without checking the window.

Fix

Add the same check from fit/row.ts to fit/column.ts. A column split is also a break at the page bottom, so it is not allowed inside a window. When the column declines, tryOverflow rewinds to the last break candidate. No candidate is recorded inside the window, so it rewinds to before the node. The node and the column then move to the next page together.

Tests

• paginate.test.ts has a new case in minPresenceAhead windows where the item after the window is a column. It also takes an image snapshot (min-presence-column-no-split.png).
• parity.test.ts has a new case in minPresenceAhead where the next sibling is a View with children. It checks that the new engine matches the legacy engine.

Both tests fail without the fix and pass with it.

I also checked a real 273-page report full of tables. Without the fix, the new engine made 270 pages and started to differ from the legacy engine at page 13. With the fix, it made 273 pages, and every page matched the legacy output pixel for pixel.

[vercel]

@exoego is attempting to deploy a commit to the diegomura Team on Vercel.

A member of the Team first needs to authorize it.

[exoego]

Sorry, will recreate