481. Mirrors, and a new wall

2026-09-09 - directed, continuing 480

The byte-order question answered itself, badly

REQ-20260909T1910Z-3d05 asked SELFish which project's doc described the other's convention. They read this crate's source and found the contradiction was inside it: NidHasher::hash said "little-endian" one screen above hash_bytes saying "Big-endian, and this was wrong for a long time", with from_be_bytes underneath. The hash wording was simply wrong.

Both docs now say it without an endian word - "packed so that digest[0] is the most significant byte" - because "little-endian" and "big-endian" describe the packing, and the packing is what differs. Each project reached for the word describing its own and called the other incorrect, in mirror-image warnings, about the same eight bytes. hash_bytes now says "wrong here" and says why the qualifier matters (D657).

SELFish corrected the same sentence on their side. Neither project's behaviour changes; what is fixed is that a crate hashes and decodes the same way round, which encode_nid's round trip pins.

The new wall

PPSA25872, at 151 imports and 321,973 calls, faults at image+0x17554a3 reading through 0xffffffffffffffff - a -1 used as a pointer. Around it:

rdi -> image+0x1e5dccc = "%s" ... "Hostname Look"
just before: libc::strlen, libc::memcpy, libkernel::scePthreadSelf

A vsnprintf/strlen/strcpy run assembling a diagnostic dump. The title has already decided something failed and is describing it, so the -1 is downstream and the thing worth finding is what put it on that path. Four calls remain unimplemented - sceUserServiceGetLoginUserIdList, sceAppContentInitialize (twice), sceAppContentTemporaryDataMount2, sceCommonDialogInitialize - and a title that cannot get a login user list has something to report. Filed as REQ-20260909T2145Z-9e52.

Surprises

• A test vector would have confirmed the wrong thing. 3d05 offered one, and it would have agreed with what both projects already knew while leaving the wrong sentence in place. Reading the source is what found it.
• Third units mix-up in two days, after the 0x6bc000 addresses and the four harness errors. Every one was two things sharing a name and not a meaning, and every one produced a confident report naming a defect somewhere.
• The trace's returned field is optional, so a field-by-field text parse of the JSON silently carries the previous call's value forward. Nearly read a shifted column as evidence; the report's own ranked findings are the reliable source and were already correct.

Next

• 9e52, and whichever of the four turns out to be the gate.
• image+0x17554a3 itself, once the path into it is known.