D562 - The tail, worked in order of what a placeholder does

Status: guest-observed Date: 2026-09-04

The question

After the event and pthread work, eighteen non-Agc functions were called and unimplemented - twenty-two calls between them, most of them once each. Ranked by call count there was nothing there. Ranked by what the placeholder does when the guest reads it, two were actively dangerous and the rest were not gaps at all.

The order that mattered

StubReturn::Unimplemented answers 0x7fff_0001. What that means to a guest depends entirely on what the function was supposed to return:

Call	What the placeholder becomes	Verdict
setjmp	non-zero - "you arrived here via longjmp"	actively wrong
setlocale	a char * the guest dereferences	D125, a wild pointer
_sigprocmask	the guest's own stack read as a signal mask	garbage in
sceKernelUuidCreate	an unwritten identifier	benign, block was zeroed
clock	a huge elapsed time	timing logic misled
the rest	a status the guest proceeds past	harmless

setjmp is the sharp one and it is worth stating plainly: a non-zero return from setjmp tells a program it got there by longjmp and sends it down an error-recovery path it was never on. The placeholder was worse than any answer, and one call was enough to justify fixing it.

Implemented: setjmp, setlocale, clock, _sigprocmask, sceKernelUuidCreate, sceVideoOutConfigureOutput.

setjmp saves nothing, and that is the record

A real setjmp stores the callee-saved registers, the stack pointer and the return address. This stores none of them, so a longjmp into the buffer would jump through uninitialised memory and take the process with it.

No title in the corpus imports longjmp - that is what makes returning zero safe today, and it is exactly the condition to check before it stops being. Implementing longjmp without first implementing the save here is the dangerous change; this one is not.

clock's unit is the biggest assumption in the batch

Microseconds, on POSIX's rule that CLOCKS_PER_SEC is 1,000,000. FreeBSD's own headers have historically defined it as 128, and the target kernel is FreeBSD-derived - so a guest compiled against those reads a time roughly 7,800 times too large. Nothing here settles which the vendor's SDK uses. First thing to change if a title's timing is wildly wrong.

sceKernelUuidCreate is deterministic on purpose

A counter, not a random value. A UUID is meant to be unpredictable and this one is not, which is a real trade taken deliberately: the only progress measure this project has is whether one run got further than the last, and an identifier that changed every run would be a difference between two traces that meant nothing - the same argument D256 makes for process time. Unique within a run, which is what a guest using one as a key needs.

What was deliberately not implemented, which is most of it

Nine of the eighteen are decisions, not gaps, and finding that out was the larger half of the work:

• sceUserServiceGetLoginUserIdList - D346 already decided this. It writes a list whose layout is unmeasured, and the entry says so where it is declared. Implementing it would have reversed a recorded decision on no new evidence.
• sceCommonDialogInitialize, sceMouseInit, sceAppContentInitialize - their modules are declaration-only by design, and the deeper reason holds anyway: reporting that a subsystem initialised when the subsystem does not exist is the exact stub principle 3 forbids. A guest told the mouse is ready calls sceMouseOpen next and gets a placeholder.
• sceVideoOutSetBufferAttribute2, sceVideoOutGetOutputStatus - fillers, not setters. Both promise to write a structure whose layout no lawful source here gives. ConfigureOutput beside them is implemented precisely because it promises nothing about the caller's memory, which is the line between the two.
• scePthreadAttrGet, scePthreadAttrGetstackaddr - D561's ambiguity, unresolved.
• PS5Util::0xa96b2b178383025c - no name, so nothing to implement.

What it changed

Nothing in the run, and that was expected: none of these were blocking, and the guest stopped at the same shader wall with the same imports. Twenty functions remain unimplemented, and every one is either out for hardware (eleven Agc and Ampr) or a decision recorded above.

What this does not establish

That any of the six is right. setlocale answers the C locale because orbistoun implements no other, not because the console does. _sigprocmask keeps a mask nothing consults. clock's unit is a coin toss between two citable conventions. Each is better than the placeholder it replaced, which is a lower bar than correct and is the bar that was cleared.