[Ernxst]

Harness: Codex
Agent: none
Model: not exposed in the recorded session metadata

Task statements

1. Fix early line wrapping caused by trailing collapsible spaces in fixed-width cells. (Working-directory and workflow instructions omitted.)
2. Defer hyphen line breaking where the current breaker needs a broader change to consider a later break opportunity before committing to an earlier space. (Working-directory and workflow instructions omitted.)

Summary

GPU-IX now lets a trailing collapsible space hang past the line box instead of counting it as overflow. The change is in the GPUI line wrapper, pinned from Zed PR #26 and Zed PR #27, at a71fa7f.

#843 remains open. On the 344 px date reproduction GPUI still wraps …Read / 2026-01-01., while Chromium wraps through …Read 2026- / 01-01.. The current line wrapper commits to the preceding space before reaching the later hyphen opportunity; fixing that needs a broader breaker change.

Verification

• With the old overflow loop, the GPUI regression test failed with left: ["aa", " ", "bb"], expected ["aa ", "bb"]. With the fix, test_trailing_space_does_not_trigger_wrap_before_next_word passes and its line strings are aa / bb.

• The balanced-wrap GPUI test also passes. Its check now asserts the contract that balanced wrapping chooses a narrower wrap width while preserving the greedy line count; hanging spaces can make the line-width spread equal even when that search works.

• bun run test in packages/react: 1,948 passed, 16 skipped.

• A scratch app with packed @gpuix/react and @gpuix/native reproduces the Watchlist row geometry: the Trigger cell is 320 px wide and 64 px tall at 12 px / 16 px. Chromium and GPU-IX both render four lines:

  The well's ~180 split is a chain of combined junctions; any
  Mk.1 pipe or junction on that path caps at 300 m³/min and
  starves the plastic leg unevenly. Any Phase 4 plastic draw
  beyond the Supercomputers would expose the shortfall.
  

  The second line ends and on both sides.

• In Chromium, Exact and Exact both fit at 29.12396 px. GPU-IX splits both at 29 px (Exac / t) and fits both at 30 px.

• At 13 px, Chromium lays out aa bb as aa / bb; GPU-IX lays it out as aa / b / b at that probe width. This exposes a separate repeated-space collapsing mismatch in white-space: normal. The GPUI change here only changes overflow accounting for a trailing space at a wrap; collapsing repeated spaces requires a broader text-layout change and is outside Trailing spaces move words to the next line in fixed-width cells #860, so this PR does not claim to fix it.

• Chromium lays out high-speed as high- / speed; GPU-IX lays it out as high / -spe / ed. This is the deferred GPU-IX does not wrap at hyphens like Chromium #843 behaviour.

• Both keep -5 on one line and wrap 2026-10-02 as 2026- / 10-02 at the tested width.

• Chromium's contenteditable case matches the four Watchlist lines; its 320 px textarea reports scrollHeight and clientHeight of 94 px. GPU-IX's 320 × 96 px textarea renders the same four lines, and its source uses the shared shape_text wrapping path.

• nowrap and pre each remain on one line. GPU-IX does not support whiteSpace: "pre-wrap", so that mode could not be compared.

• No screenshot goldens changed in the test run.

Closes #860