Use ratio-based perf checks, not absolute ms, on shared CI hosts

Absolute timing assertions (e.g. 'placed 2000 marks in under 500 ms') flaked between 537 and 1800 ms when machine load ran 90-600. Those checks went red while the code was fine. A ratio check stays stable: time the big case against a small baseline in the same run, take the best of 3 for each, and assert big < k x small (+ a few ms of slack). It still catches the O(n^2) regressions you care about (e.g. 17.9 s vs 149 ms before a fix, 127 ms vs 12 ms after). Similarly, replace fixed sleeps in UI integration tests with polls that have a deadline; fixed sleeps were the cause of most load-related flakes.