Problem

• WebNN has no performance surface: no timing, no attribution.
• dispatch() has no completion signal: "callers can await the results of reading back the
  output tensors."
• The only way for a page to learn that a run finished is a readback — so the readback becomes
  part of the number, and the bracket differs per engine. Page-timed numbers are not interoperable.

Scenario

A page runs a face detector and wants to pick a smaller model on a slow device. It cannot tell
whether the slow device is limited by the model, the hardware, or its own input preparation.
The only measurement it can take is this:

const start = performance.now();
context.dispatch(graph, inputs, outputs);
await context.readTensor(a);          // the only completion signal
await context.readTensor(b);
console.log(`Inference: ${(performance.now() - start).toFixed(2)}ms`);

On a reference mobile device (NPU backend):

The page's "inference" is 209 ms against 5.7 ms of computation, and it cannot see the difference.

Proposal

MLComputeTiming: a PerformanceEntry interface with two engine-created entry types.

[Exposed=(Window,Worker)]
interface MLComputeTiming : PerformanceEntry {
  readonly attribute DOMHighResTimeStamp computeDuration;
  readonly attribute boolean accelerated;
  [Default] object toJSON();
};

• One entry per execution, never per operator.
• Observer-only: availableFromTimeline: false, delivered through PerformanceObserver.
• duration = submission → completion (queueing included); computeDuration = the engine's own
  compute interval inside it.
• graphId (engine-assigned, per context) correlates compile and run entries; accelerated is
  one bit.

Relations

• #934 (effective compute policy):
  complementary — this contributes measurements and the privacy framing, not a device surface.
• #923 (compute policy rename):
  decides the long-term shape of the accelerated field.
• Model Loader (draft): the same entries are
  designed to cover its load() / compute() once that API stabilises.
• WebPerf: the entry types need a row in the timing entry types registry. WebNN owns the
  semantics; WebPerf owns the registration.

Measurements, alternatives, privacy and the open questions (naming, graphId vs MLGraph.id,
accelerated shape, computeDuration normativity) are in the explainer: link when published.