[danwalmsley]

Dragging Kestrel’s properties divider updates the root --right-width variable, but currently rebuilds styles throughout the root subtree. Limit recalculation to the mutation root and dimension consumers when dependencies are safe. Preserve style batching and synchronous CSSOM barriers; mixed mutations, aliases, inherited properties, pseudo-elements, shadow DOM, and style-attribute dependencies retain the full path.

Explicit dimension inheritance also retains ancestor-first subtree recalculation, including var() fallbacks resolving to inherit. Replay inline width/height inheritance during a full cascade so descendants follow their parent while preserving stylesheet priority.

The full-subtree setter exists on main and dates to commit 869f1a0 (July 20), predating the recent Windows work. This identifies a longstanding invalidation bottleneck; it does not claim a complete historical frame-rate bisect.

After the review correction, three alternating Windows runs of the included 1,000-item DOM benchmark measured 29.92 ms baseline versus 3.17 ms fixed median synchronous resize time (about 89% less). This measures style/layout, not full Kestrel FPS or physical presentation.

Validation:

• Reproduced the review regression before fixing it.
• Added 24 inheritance combinations covering stylesheet/inline declarations, literal/var-fallback inheritance, one/three descendant levels, and setProperty/removeProperty/empty setter, plus an inline-priority case.
• All 14 focused compatibility checks and all three parser suites pass on the corrected patch. The baseline control independently confirms an older literal-inline inheritance issue fixed here as well.
• The full native suite fails on unchanged main and the patch at the existing iframe preparation did not overlap discovery with outer script work assertion; that unrelated test remains unchanged.

Commands, history, and raw results are in docs/validation/css-variable-resize.md. This PR is based directly on main and includes no Windows WebGPU, pacing, font-cache, or Kestrel fixture changes.

[danwalmsley]

The optimization looks beneficial and complements #50, but I found a correctness regression that should be fixed before merging.

At cbd32569, the consumer selection in recascade_dimension_custom_property only includes direct variable consumers. It misses descendants that inherit the resulting dimension:

<style>
  :root { --size: 100px; }
  .parent { width: var(--size); height: var(--size); }
  .child { width: inherit; height: inherit; }
</style>
<div class="parent" id="parent"><div class="child" id="child"></div></div>

After an initial geometry read, run:

document.documentElement.style.setProperty('--size', '200px');

Observed [parent width, child width, child height]:

• Baseline: [100, 100, 100] → [200, 200, 200].
• PR Limit dimension-only CSS variable recalculation to affected elements #51: [100, 100, 100] → [200, 100, 100].

Both width: inherit and height: inherit need dependency propagation. Please recascade indirect dependents in ancestor-first order, or fall back to the affected subtree when those dependencies cannot be tracked safely. The same rule should account for other supported dimension dependencies, rather than assuming that a non-inherited-by-default property cannot affect descendant styles.

Please add isolated regressions for stylesheet and inline inheritance, multiple inheritance levels, and variable updates through setProperty, removeProperty, and the empty-value setter. The existing dimension-variable-compatibility filter passed locally despite this regression.

Performance evidence is encouraging: one local macOS run of the included 1,000-item benchmark measured a median of 17.62 ms on the baseline versus 1.56 ms with this PR, across 60 samples after warm-up. This is a single synchronous style/layout microbenchmark comparison, not an application FPS measurement.