[FanWu-ai]

max_layout_dimensions treats every zero-preferred dimension as empty, discarding its finite maximum. Wrapping Window(width=Dimension(max=10, preferred=0)) in an HSplit therefore removes its 10-column cap: an adjacent VSplit assigns it 40 columns out of 80 instead of 10.

Ignore only dimensions whose maximum is zero. Preserve explicit finite caps while retaining the existing rule that other children's larger preferred sizes take precedence. Invisible children still cannot shrink visible siblings.

Add eleven tests covering a lone cap, default/invisible siblings, multiple caps, positive preferred siblings, empty inputs and a real nested split. Eight regressions fail on unchanged source.

Validation on Linux / Python 3.12.14:

• Fresh pristine complete suite: 156 passed; final complete suite: 167 passed, no skips.
• Ruff lint/format, typos, and strict Mypy checks for Linux, Windows and macOS pass. Those are platform-aware static checks, not native OS test runs.
• All twelve configured pinned hooks, wheel/sdist builds, strict Twine checks and git diff --check pass.
• Independent automated review reran the full suite and required static checks, reproduced the old failures, and passed 7,395 additional assertions including both real layout axes.

Other interpreter/native OS and hosted CI results are not claimed. Prepared and tested with OpenAI Codex assistance; no human review is claimed.