An eboot is not the module this project builds, in three separate ways

/app0 mounts, and a console's rtld now reads the eboot and refuses it. Each refusal named the next difference, and all three are the same kind of difference: what a loader that maps a module wants, against what a loader that executes one does.

	module (elfldr, emulators)	eboot in a real package
container magic	current generation	previous - and selfish wrap already defaulted to it
e_type	0xFE10 executable	0xFE00 fixed-address executable
image base	0 - addresses are offsets, the loader adds a base	0x400000 - mapped where the file says
first PT_LOAD file offset	0, with FILEHDR PHDRS	0x4000, and the headers are in no segment at all

The first two are fixed: mkmodule --fixed writes 0xFE00, the eboot targets wrap at generation 4, and module.ld takes an optional OBS_IMAGE_BASE.

The fourth row is the one still standing, and it is the interesting one because this script argues for the opposite in a comment that is correct about the case it was written for:

FILEHDR and PHDRS here rather than on a segment of their own: without them the headers are not covered by any segment, and a loader that maps only what the segments describe cannot read the header table it just used.

True of a loader that maps segments and then reads the file through the mapping. A system loader reads the file directly, and refuses one whose first segment starts at file offset zero:

[rtld] ERROR verify_ehdr:930: B: offset 0x0  end 0x190

0x190 is exactly ehdr + 6 phdrs. A real eboot puts its first LOAD at file offset 0x4000 with PT_INTERP at the same place, so the header region is covered by nothing. Both layouts are right for their loader, which is why this cannot be a fix to module.ld - it needs a second script, or a switch as deliberate as OBS_IMAGE_BASE.

Worth saying about the method

Every one of these came from measuring a real package's eboot rather than from reasoning about what a loader ought to want, and three times the reasoning would have been wrong: a current-generation console wants the previous generation's container, a fixed-address executable is a different e_type rather than a linker flag, and mapping the headers - which one loader requires - is what another refuses. The file was sitting inside a package that has been on the bench all along.