Summary

Fixes #21916

When a JSX expression child whose flat width fits at its current indent is followed by literal text such as (, the formatter was breaking the expression across three lines ({, content, }) even though the line as written is within printWidth. Prettier leaves it inline.

Input

const App = () => {
  return (
    <O>
      <I>
        <x />
        {f(xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx).y} (
        {label})
      </I>
    </O>
  );
};

Before (incorrect)

const App = () => {
  return (
    <O>
      <I>
        <x />
        {
          f(xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx).y
        } (
        {label})
      </I>
    </O>
  );
};

After (matches Prettier)

const App = () => {
  return (
    <O>
      <I>
        <x />
        {f(xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx).y} (
        {label})
      </I>
    </O>
  );
};

Root cause

In the printer's Fill, each entry's fit is measured with a flat-only FitsMeasurer bounded by SingleEntryPredicate, so the fill correctly determined the expression entry fit on a single line. While printing that flat entry, however, the inner expression group (group(["{", indent([softline, expr]), softline, lineSuffixBoundary, "}"])) re-ran its own fits check via Printer::fits, which uses AllPredicate and walks past the entry boundary into subsequent fill entries. The accumulated width frequently exceeded printWidth, so the inner group expanded — even though the fill had already proven the entry fits flat at the current column.

Soft line breaks don't terminate a flat-mode fits walk (they're a no-op), so the measurement happily kept reading through the next item, separator, and so on until either a hard break or the end of the document.

Fix

Adds an in_flat_fill_item flag on the printer state that's set while print_fill_item is printing an entry in flat mode. The StartGroup arm honors it and treats nested groups the same way it would when a parent group has already verified the fit — i.e. it skips the re-measurement and prints the inner group flat.

The flag is intentionally scoped to regular groups: BestFitting still goes through print_best_fitting's variant selection so JSX child lists with both flat and expanded variants (the <Link>BREAK THIS</Link> pattern in issue-16199.jsx) continue to choose the multi-line variant when the surrounding context is too wide for the flat one.

Prettier conformance: no regressions (746/753 JS, 591/601 TS — unchanged).

AI Disclosure

This PR was co-authored with Claude Code (AI assistant), as noted in the commit. The fix was reviewed, tested against the full Prettier conformance suite, and verified to produce no regressions.