[nicoburns]

Summary

On main, the browser-layout-bench real/wikipedia leaf_10pct relayout step (every 10th leaf +1px, one incremental resolve()) gives boxes that differ from a fresh layout of the serialised DOM (775 differing elements). Two independent layout-cache bugs cause it:

1. Same-BFC children cached without their float context. LayoutBlockContainer::compute_block_child_layout wrapped same-BFC child layout in compute_cached_layout, but the mutable BlockContext (floats placed so far, and floats the child itself contributes) isn't part of LayoutInput/the cache key. A cache hit therefore skipped replaying the child's floats into the shared context and reused a layout computed against a different float environment. E.g. <div><div style="float:right;height:200px"></div></div><div style="overflow:hidden">…</div>: changing the second div's width left the float wrapper cached, the float was never re-placed and body came out 0px tall instead of 200px.
   Fix: bypass (and clear) the cache for a same-BFC child when the context can affect it or it can affect the context:

   let affected = if is_inline_root { ctx.has_active_floats(0.0) } else { ctx.has_floats() };
   if affected || self.subtree_contains_floats(node) { clear_layout_cache; compute uncached }

   subtree_contains_floats is memoised per layout pass and is O(1) for inline roots without inline boxes (floats inside an inline formatting context are inline boxes). Block children keep the conservative has_floats() because negative margins can bring them above their parent.

2. Atomic inline measurements lost baselines and moved children. Inline layout sized atomic inlines with the caller's run_mode, so under ComputeSize Taffy's cache returned LayoutOutput::from_outer_size (no baselines) on a hit, while a recompute had baselines → different line geometry between incremental and fresh layout. Also the inline-box placement loop ran during ComputeSize, overwriting child positions that a later parent PerformLayout cache hit then didn't restore (wikipedia navbox th inline-block at y=5 vs 17).
   Fix: atomic inline children are always laid out with RunMode::PerformLayout (baselines needed), and inline-box positions are only written during PerformLayout.

Tests

incremental_oracle.rs: cached_same_bfc_child_replays_floats, cached_inline_root_moves_below_floats, cached_atomic_inline_measurement_preserves_baselines — each fails without its fix. blitz-runner --check on wikipedia, text/paragraphs/1000, flex/wide/1000: every step matches full restyle and fresh parse (rustdoc's pre-existing round-trip difference is unchanged).

Performance

main ebacd68 vs this PR 9f8d36b, separate release builds, pinned to one core, 10 interleaved process pairs (ABAB) × 5 iters = 50 samples per cell. Layout phase of the timed resolve(), ms, median [p25–p75]; bold = IQRs don't overlap.

case	step	main	PR	Δ median
real/wikipedia	initial	18.5 [17–20]	20 [18–21]	+8%
	root_width	8.8 [8.6–9.8]	9.5 [9.3–10]	+8%
	leaf_1	4.3 [4.1–4.7]	6.3 [6.1–6.7]	+47%
	leaf_10pct	7.7 [7.5–8.5]	9.6 [9.4–10]	+26%
	leaf_all	17 [16–19]	17 [17–19]	0%
	append	20 [18–22]	21 [19–22]	+5%
text/paragraphs/1000	initial	7.1 [6.9–7.8]	7.3 [6.9–7.9]	+2%
	root_width	7.0 [6.8–7.3]	7.0 [6.8–7.1]	0%
	leaf_1	0.232 [0.229–0.240]	0.290 [0.284–0.295]	+25%
	leaf_10pct	0.95 [0.94–1.0]	1.0 [1.0–1.1]	+6%
	leaf_all	7.1 [6.8–7.4]	7.1 [6.9–7.5]	−1%
	append	1.1 [1.0–1.2]	1.15 [1.1–1.3]	+5%

The cost is in small incremental relayouts: about +2 ms on wikipedia leaf_1 (+42% of total resolve(), 4.8 → 6.8 ms) and about +2 ms layout on leaf_10pct (+3% of total resolve(), which is dominated by style); paragraphs leaf_1 is about +60 µs. Other steps are within noise. Most likely this comes from the conservative cache bypass (float-adjacent same-BFC subtrees laid out uncached, plus the per-pass float-subtree walk). The phase timer rounds ≥10 ms values to whole ms.

Link to Devin session: https://dioxus.staging.devinenterprise.com/sessions/48e12aa27062474cb1529350310c9c58
Open in Devin Desktop: https://dioxus.staging.devinenterprise.com/desktop/session/48e12aa27062474cb1529350310c9c58?variant=devin-insiders

[staging-devin-ai-integration]

I'll fix CI failures and address comments from users with write access that start with 'Devin'.

• Disable automatic comment, CI, and merge conflict monitoring