Thanks to Marlin pressing me on these loader glitches, we finally figured out the reason: the high-speed OS. The patched OS is really unnecessary now that Hias' high-speed SIO driver is built into the PBI BIOS, but since Candle left the high-speed OS in the default OS slot, a lot of people end up running it anyway. Whether they inadvertently end up running it alongside the PBI implementation (which I imagine would cause wildly unpredictable results) I do not know, but simply running the latest XEX loader with the high-speed OS revealed a problem caused by the patched OS replacing the loader's VBI code with its own during SIO calls. This frequently resulted in the loader's internal DLI counter going out of sync (since the VBI is responsible for resetting it at the top of each frame), causing a crash.
The PBI implementation of Hias' driver does not cause this problem, since it only subverts the currently running VBI if the currently active SIO operation requires it (i.e. a divisor 0 transfer). The loader performs no high-speed IO whatsoever (indeed, it only issues API calls to the PBI BIOS), so there was zero chance of problems there. Unfortunately the patched OS relies on certain decisions being made at compile time, and the version of the patched OS in circulation clearly runs its own minimal VBI handler during SIO calls.
In any case, and despite the fact there's little or no need to ever use the high-speed patched OS with the loader, the fix was easy enough and will be part of the final firmware. Hopefully that puts an end to this mystery once and for all.