Don't let a deferred float inflate its block formatting context

[mvanherwijnen]

What's wrong

When a block that establishes a block formatting context (display: flow-root, overflow other than visible, etc.) contains floats, a float that doesn't fit on the current page can push the entire block to the next page

This is very visible in print documents full of figures that use float + wrapping text: blocks jump to a fresh page instead of partially filling the current one.

Minimal example

<style>
  @page { size: 100px; margin: 0 }
  body { margin: 0 }
  div.bfc { display: flow-root }
  p { height: 20px; margin: 0 }
  div.float { float: right; width: 30px; height: 40px }
</style>
<div style="height: 70px"></div>
<div class="bfc">
  <p>a</p>
  <div class="float"></div>
  <p>b</p>
</div>

• Expected: page 1 = spacer + a; page 2 = float + b.
• Actual: page 1 = spacer only; the whole .bfc (both paragraphs and the float) is pushed to page 2, wasting ~30px of page 1.

Cause

float_layout appends every laid-out float to context.excluded_shapes. When the float overflows and _out_of_flow_layout breaks before it (deferring it to the next page via resume_at), the float is left in excluded_shapes.

finish_block_formatting_context then grows the BFC box's height to max(shape bottoms) including the deferred float, which floated up beside the content that stays on the page. The fragment's height therefore exceeds the page, the caller rejects it, and the whole block is moved to the next page instead of keeping the part that fits.

Fix

When breaking before an overflowing float, drop it from the current formatting context's excluded shapes so it no longer inflates the fragment that stays on the page:

context.excluded_shapes = [
    shape for shape in context.excluded_shapes
    if shape is not new_child]

Test

Adds test_float_overflow_splits_formatting_context in tests/layout/test_float.py, reproducing the example above and asserting the block is split (paragraph a stays on page 1). It fails before the change and passes after. The full tests/layout/ suite passes with the change.

[liZe]

Hi!

Thanks for the pull request.

The problem is actually larger than what your patch fix. I’ll add a commit on top of yours to fix that!