perf(binding): cache identical measure requests

[hyfdev]

Summary

• cache successful identical measure results inside the Rust MeasureSession created for each computeLayoutWithMeasure()
• key requests by raw node ID, both optional known dimensions, and both available-space variants and definite f32 bit patterns
• reuse a hit before creating callback arguments, converting Style, or crossing Node-API/WASI
• cover exact reuse, every key component, per-compute lifetime, thrown-value identity, failure retry, and shared Native/WASI behavior

Root cause

Taffy can issue the same measure request repeatedly during one layout computation. The binding rebuilt MeasureArguments, converted the complete Style snapshot, crossed Node-API or WASI, and called JavaScript for every repeat even though the node and all constraints were identical.

In the coding-agent chat initial-layout scenario, 3,420 Taffy requests collapse to 1,262 exact request combinations.

Implementation

Each raw measured compute already creates a fresh MeasureSession. The session now owns a HashMap<MeasureCacheKey, Size<f32>>, so its entries are dropped automatically when that compute returns.

The key contains:

• raw node ID
• known width and height, including None versus exact f32::to_bits()
• available width and height, distinguishing Definite(bits), MinContent, and MaxContent

The failure check remains first, so a failed session continues to drain later Taffy requests with zero sizes. A result is inserted only after the JavaScript call and conversion to Size<f32> both succeed. Callback throws and binding conversion failures are not cached; subtree invalidation and original thrown-value identity remain unchanged.

No cache survives into another compute, and this does not add Style snapshots, dirtying rules, algorithm changes, or a public API.

Performance

Measured on Node v24.19.0, Linux x64, Intel Core i5-13500H. Each latency is the median of three independent local benchmark runs.

Package	Before	After	Change
@taffyjs/node	44.2894 ms	18.1828 ms	-58.9%
@taffyjs/wasm	356.2007 ms	133.2282 ms	-62.6%

The public measure callback count changed from 3,420 to 1,262 for both Native and WASI. Temporary counting code was removed before commit.

A complete local benchmark run showed no obvious regression in the three scenarios without measurement callbacks. The measured viewport-resize scenario also improved from 19.0808 ms to 8.8367 ms for Native and from 158.6806 ms to 65.7667 ms for WASI relative to the same-machine published baseline.

Validation

• vp run check — passed, including format, type-aware lint, Clippy, 10 Rust tests, 204 Native public behavior tests, native binding tests, type tests, and Yoga tests
• vp run check:wasm — passed, including the same 204 public behavior tests on WASI, browser runtime/bundle checks, type/package checks, packed consumers, Yoga WASI, and website build
• vp run benchmark --scenario=coding-agent-chat-initial-layout — three runs before and three runs after; medians reported above
• vp run benchmark — complete local suite passed with equivalent layout validation
• bounded independent adversarial review — PASS

Fixes #38

[Copilot]

Copilot encountered an error and was unable to review this pull request. You can try again by re-requesting a review.