[hyfdev]

Measure callbacks receive the same public MeasureArgs values. The private crossing now carries the constraints in a caller-owned reusable Float64Array record, and the callback argument is only the existing getStyle handle. Extra properties on the returned size are ignored.

tree.computeLayout({
  root,
  availableSpace: { width: 400, height: 800 },
  measure({ knownDimensions, availableSpace, node, getStyle }) {
    return { width: 40, height: 16 };
  },
});

A return of { width: 40, height: 16, extra: 3 } measures as { width: 40, height: 16 }. Missing fields, null, arrays, and non-number payloads still throw TypeError.

Why

Each measured layout pauses into JavaScript once per exact request. The cost of that pause was dominated by building a nested constraint object per request and enumerating the returned object's keys. In the coding-agent chat scenarios, #computeMeasuredLayout itself was 68.9% and 77.6% of CPU on the @taffyjs/node path.

The record keeps the private transport independent of the public representation, so no numeric slot has to double as its own tag:

• Presence of each known dimension and the AvailableSpaceKind of each axis occupy one tag slot. A caller-supplied definite -1 or -2 therefore stays a definite available space rather than becoming MinContent or MaxContent.
• Taffy's node id occupies two slots as 32-bit halves. Its slot version advances on every reuse, so ordinary node churn walks the id past the range a double holds exactly; split halves keep every id addressing the measure function and context that belong to it.
• The wrapper rebuilds knownDimensions, availableSpace, and NodeId before calling the user function, and those values stay ordinary mutable objects created per request.

One shared record serves the common case. A measure callback may lay out another tree, so a nested compute takes its own record and leaves the outer one untouched.

This does not change how often Taffy asks for a measurement, and it does not make @taffyjs/node match yoga-layout's callback count.

Performance

Same machine, Node v24.19.0, local vp run benchmark profile, main and this branch measured in alternating rounds. Scenarios were not edited.

Scenario, @taffyjs/node	main	This branch
coding-agent-chat-initial-layout	3.5746 ms	2.2199 ms
coding-agent-chat-viewport-resize	1.4675 ms	0.9132 ms

The yoga-layout control moved by less than the spread between rounds: 0.8129 ms and 0.9057 ms for the initial layout, 0.2078 ms and 0.2085 ms for the resize.

benchmarks/results/published.json is unchanged.

Validation

• vp run check
• vp run check:wasm
• A definite -1 and -2 available space reaches the callback as Definite, alongside the existing -4, -0, NaN, and infinity cases
• A node id past the exact-double range still reaches its own measure function and context
• Extra-field measure return succeeds without changing the measured size

[hyfdev]

Landed on main as 8e9d6ce.

That commit carries the measure constraints in a caller-owned reusable Float64Array record rather than in named fields on the callback argument, and the callback argument is only the getStyle handle. Every numeric slot then carries only its value: presence of each known dimension and the AvailableSpaceKind of each axis occupy one tag slot, so a caller-supplied definite -1 or -2 stays a definite available space, and Taffy's node id occupies two slots as 32-bit halves, so an id whose slot version has advanced past the range a double holds exactly still reaches its own measure function and context.

Ignoring extra own properties on a measure result lands unchanged.