Performance measurements taken in a hidden browser tab are worthless — and they look like a freeze

Symptom

You are chasing a report of "the browser freezes on this page". You drive the page from the console (or from an automation tool that evaluates JavaScript in the tab) with something like:

for (let i = 0; i < 40; i++) { scroller.scrollTop += 200; await new Promise(r => requestAnimationFrame(r)); }

The call never returns. The tool reports a renderer timeout. It certainly looks like you reproduced the freeze — the page stopped responding to your script, which is what "frozen" means.

Switching to timers does not help:

for (let i = 0; i < 25; i++) { scroller.scrollTop += 200; await new Promise(r => setTimeout(r, 16)); }

That, too, times out.

Cause

The tab was in the background. Browsers aggressively suspend hidden tabs:

• requestAnimationFrame callbacks do not fire at all while a tab is hidden. There are no frames to paint, so awaiting one waits forever.
• setTimeout / setInterval are clamped to roughly one second in background tabs (and can be frozen entirely after several minutes of inactivity). Your 25 × 16 ms loop is not 0.4 s; it is 25 s or worse — long enough to blow past a typical automation timeout.

Both failure modes surface as "the renderer may be frozen or unresponsive", which is the same message you would get from a genuine runaway loop.

There is a second, quieter half to this, and it is the one that actually invalidates the experiment: the code you are testing is throttled too. Any layer that schedules its work through requestAnimationFrame — most redraw and layout-sync code does — simply does not execute while the tab is hidden. So a hidden-tab run can report a beautifully flat memory graph and zero observer wake-ups, and prove nothing whatsoever.

Fix

Check visibility first, and treat a hidden tab as "no result", not as a result.

document.visibilityState   // 'visible' | 'hidden'
document.hidden            // true = anything you measure next is suspect

Then:

• Run interactive/timing measurements only in a foreground tab. If your automation cannot guarantee focus, say so in the write-up rather than reporting the numbers.

• In a hidden tab, measure synchronously. One block, no await. Reading layout (getComputedStyle, offsetHeight) forces a synchronous reflow, so you can still measure the cost of a DOM operation; you just cannot measure anything that depends on frames or timers elapsing.

• Layout reads are trustworthy — transitioned layout reads are not. Forcing a reflow gives you real numbers for anything you set directly, but a property under a CSS transition or animation needs frames to move, and a hidden tab paints none. getBoundingClientRect() on such an element keeps returning the geometry from before your change, indefinitely, while element.style.top already holds the new value. Read the inline style (or the model behind it) rather than the box, or drop the transition for the duration of the measurement:

  el.style.transition = 'none';   // then the box and the style agree again

  This one is nastier than the timer clamp, because the numbers look plausible: every measurement returns the previous step's position, so a walkthrough of N steps reports N sane-looking rectangles that are all off by one.

• Treat zeros with suspicion. In a hidden tab, "0 redraws, 0 observer wake-ups, flat heap" is more likely to mean suspended than healthy. A negative result from a suspended tab is not evidence of absence.

• Prefer in-page instrumentation over remote driving for anything that must run over time: have the page count events into a global itself, then read that global later in a single synchronous call. The counters keep working across your own tool's round-trips, and the numbers come from whatever the tab actually did.

Why this is worth a page of its own

The failure is self-confirming in the worst way: you are looking for a freeze, and the environment hands you something indistinguishable from a freeze. Without the visibility check there is nothing in the output to tell you the measurement was void, so the natural next step is to report a reproduction that never happened.