A flaky timing test no longer fails the build

A build on main failed on a performance test that checks the server stays responsive while many cards land at once. The code was fine: the test timed itself with the wall clock, and on a busy CI runner the process spent part of that time waiting for a CPU core. The test now measures the CPU time the server actually spends, which stays steady under load. As a side effect, it now also catches the slowdown it guards against on fast machines, where the old timing let it pass.