Project: permit browser regression investigation

Build contract

• Collect supported browser timing entries, record missing types, cap resource events, and treat telemetry transport failure as nonblocking.
• Mark one approval interaction and divide latency into input delay, handler work, and presentation delay under both idle and competing work.
• Run comparable photo-panel lifecycle cycles and inspect a retained path for listeners, observers, object URLs, or callback registries.
• Compare alternating DOM write-read updates against a batched version, then check overlays, font loading, zoom, and keyboard focus.

Implementation checkpoint

function chooseInvestigation(inputDelayMs, handlerMs, presentationMs) {
  const phases = [["input-delay", inputDelayMs], ["handler", handlerMs], ["presentation", presentationMs]];
  phases.sort((left, right) => right[1] - left[1]);
  return phases[0][0];
}
console.log(chooseInvestigation(190, 14, 106));
// Output: input-delay

Cost and boundaries

Limit performance events by a documented per-view cap and sampling rate; collection itself consumes main-thread and network budget. A worker or chunked filter can reduce input delay but has scheduling and data-transfer cost, while immediate approval feedback needs a rollback state if the server rejects it. Memory snapshots pause or perturb the page, so use equivalent cycles and inspect retaining paths rather than comparing arbitrary heap totals. Batching row geometry avoids repeated layout, but stale widths after a font or viewport change can produce wrong placement. Paint containment may clip the photo-panel menu. Evaluate total user-visible response, not only a shorter JavaScript function.

Failure drill

Disable a timing entry type and verify its absence is recorded rather than logged as zero. Block telemetry upload without preventing an approval. Start the 47,000-record filter just before the approval click, then show whether queueing or handler time dominates. Reject the approval from the server and check that optimistic feedback returns to an accurate state. Open and close the photo panel 47 times, compare retained panel objects after similar collection points, and inspect one retaining root. Finally, trigger 47 badge updates on 620 rows, test a batched measurement, and open the anchored menu at a zoomed viewport to catch accidental clipping.

Acceptance checks

• Each field percentile includes its route cohort, release, sample count, and missing-measurement count.
• The slow interaction is attributed to a measured phase and retested under competing work.
• Panel resources release on close and late callbacks cannot revive the removed view.
• The table update reduces repeated layout without clipping controls or losing keyboard access.

Common Mistakes

• Calling a metric zero when its browser entry type was unavailable.
• Optimizing only the approval handler when another task delays its start.
• Comparing unrelated heap snapshots or clipping a menu with containment.

Related lessons

Browser Performance Diagnosis and Measurement; Field Performance Observation and Sample Contracts; Interaction Latency Traces and Main-Thread Contention; Retained DOM Memory and Lifecycle Investigation; Layout, Paint Regressions, and Containment Decisions.