The problem

Interaction to Next Paint (INP, the time from a tap to the next frame the browser draws) failed Google's 200 ms threshold on category pages for phone users in every market I looked at. The slow taps clustered right after a filter change.

What I found

A filter change re-rendered the whole grid of 48 cards in a single pass that the browser could not interrupt. React 18 only splits a render into slices when the update is marked as a transition. Here the update arrived from the data library through a synchronous subscription, so React treated it as urgent and rendered all 48 cards before letting the browser respond to anything.

A second cost hid in the same frame: the code that decides which cards are near the viewport read the layout on every pass, which forced the browser to recalculate it.

What I changed

I memoised the value the grid reads and passed it through React's deferred value, so the urgent pass finishes immediately and the heavy render happens as a transition the browser can interrupt for a tap. The navigation flag got the same treatment, and the layout read became a fixed margin.

I also dropped one planned change. Moving the page-view event after the paint looked sensible, but measurement showed it already ran in its own task, so there was nothing to gain.

The result

In the lab, on a production build with a 4× CPU slowdown, median tap time during the re-render went from 56–88 ms to 40–56 ms, and React's share of each tap fell from 62–112 ms to 12–25 ms.

In the field, the category page's INP for phone users fell from 227 to 196 ms in the US, 295 to 228 ms in the UK and 317 to 233 ms in Germany within a month. Several fixes shipped in that month, so the field number belongs to all of them together, not to this change alone.