Is it database requests? RPCs? Is there a function call which is computationally expensive?
Teams then respond by making rules to help make slow things stand out "no RPCs in functions, no synchronous RPCs, only 1 database read per request, don't call this code in a loop", etc.
Caught before merge? Only if you have someone really familiar with all the abstractions _and_ the team's previous failures reviewing it. Otherwise, developers are always finding new ways to write slow code. :)
But, that's why we have feature flags and incremental launches.