The same benchmark against 1000, 5000 and 20000 files, changing only the fixture: 12-14 frames, 41 frames, 184 frames. That is roughly linear, which is what a RecyclerView fed a pre-sorted list should do, and it is the shape to keep watching - a superlinear jump would mean per-row work that belongs to the folder.
The number worth keeping an eye on is not the median but the maximum at 20000: 598 frames against a median of 184, so one iteration of five stalled at roughly three times the others. At that size the first open is occasionally not just slow but stuttering, which is exactly the case a release should be checked against.
PERFORMANCE.md records the cost of the fixture too, because it dwarfs the measurement: 20000 files took about thirty minutes through MIUI's storage layer against four minutes of benchmark, which is an argument for building the fixture once and keeping it. And it records where the jank metric came from, since this phone cannot produce one: the API 35 emulator at the 1000-file point gave frameDurationCpuMs P50 218 ms and frameOverrunMs P50 315 ms, software-rendered and therefore not comparable, but proof that the metric exists.
SoftwareApp
65569e9af2