The only difference I can see between the 2 DLLs is the base address, as mentioned before. If you use MS's rebase, and set the base address in broken DLL to 0x62F00000 (base address of the good one) what happens? Have you tried using Dependency Walker on the broken DLL? IDK what pmver is. What is the difference between the 2 DLLs in your experience, compiler flag wise? Since the only different is the base address, it sounds like there is a collision with something that isn't relocatable, or the DLL isn't relocatable for some unknown reason (unknown reloc type entry for example) even though it has reloc info. Also, Dynaloader uses perlhost.h/vdir to possibly change the path, so its not always the path that Dynaloader is reporting exactly. I know you said if you change the file the not a win32 application happens. If you look at http://support.microsoft.com/kb/113996
you will notice many different NTSTATUS codes map to ERROR_BAD_EXE_FORMAT. Using a C debugger to look at the EAX return value of LdrLoadDll in LoadLibraryExW. That will be the native status code. To figure out if relocation failed, set a breakpoint right before "/DynaLoader.pm line 226" (I use system("pause");), run the perl process until the PLPlot load attemp, then use VMMap or MemMapView and look at the address space, and see if there is anything at or near 622c0000 to 0x625D7200 (last section 0x625ce000 + last section length 0x9200 = 0x625D7200 ), then hit space to unfreeze it, and see if it fails with not a valid application.
edit:When boot of PLplot is called, PLPlot's DLL is fully loaded, and the LoadLibrary returned a while ago. "Can't load PDL::Core module" is a croak call in PLplot's boot, so using Dynaloader for broken and ok DLLS, in isolation, is fine. But the PLplot PM file use/load/requires something else, this something else is causing the problem that later causes the bootstrap/LoadLibrary to fail for broken PLplot dll.