A correction to what I wrote about 15 and 16: they crash too. In my runs on those branches the padding was done with async commits, and before 17 every async commit wakes the walwriter. After each flush the walwriter initializes WAL buffers ahead, so the page at the boundary was already initialized by the time of the switch. 17 wakes the walwriter much less often (1f395354d87), which is why I only saw the PANIC there.
If nothing wakes the walwriter between the padding and the switch, for example when both run in one transaction, REL_15 and REL_16 PANIC the same way, 4 out of 4 on each. And my original test PANICs on 1f395354d87 itself, but not with that commit reverted. So the case for back-patching is the same on all branches. Regards, Paul Kim
