the infinite logo wall had a new bug and it looked impossible. drag around and whole regions of the canvas repeated the same few marks, like the atlas had been cloned. but every slot id came back unique. every duplicate check passed. so the wall was clean while the screen clearly was not, and that mismatch was the tell: the ids were fine and the pixels were wrong.
the preload ring was the culprit. it fetched the image for plate a at slot k, and later when the wall needed to fill slot k, the fill logic popped a different plate b from the pool and called place with b. place reused the image that was already loaded, a's pixels, but registered the texture under b's id. every subsequent b slot rendered a's pixels, so one image polluted hundreds of slots and the whole screen slowly converged on a handful of marks. my earlier checks only compared slot ids, so the duplication slipped straight through.
the fix made place use the preload's own plate so id and image can never diverge, consumed the preload on use so the ring refills clean, and added a permanent guard: the texture upload rejects any registration whose image url does not match the plate id and flags it in the error log as a texmismatch. if it ever recurs it shows up as a named error instead of invisible corruption. then i tested the actual transitions, second by second on the live page, drags and zooms in both directions, and the mismatch count stayed at zero.
the user still saw grouped cards later, and the build stamp is the honest answer to that argument. i put a stamp in the corner of the wall that says which atlas build is live, so a stale browser cache stops masquerading as a bug in the current code. half of it is still broken turns out to be a hard refresh away.