2026-09-03 - (/loop) The vendor `stat` is not the POSIX one under another name

181 -> 188 distinct imports (+7)   415,356 -> 415,532 calls
suites 125   tests 2011   clippy/fmt/identity clean
wall unchanged: read of 0xa0 at image+0x1389269

Eleventh cron tick. sceKernelStat was declared at arity 2 and registered to nothing, while stat - the same call, POSIX-named - had been implemented all along.

They agree about success and disagree about failure

POSIX  stat           answers -1 on failure
sce*   sceKernelStat  answers 0x8002_00xx, as sceKernelMkdir beside it already does

Registering the POSIX function under the vendor name hands a caller 0xffff_ffff_ffff_ffff where it tests a negative 32-bit vendor error. Not the same number, and invisible until a guest branches on it (D125). The success path and the structure written are shared, so which generation of struct stat a guest gets stays one decision (D374).

ENOENT because that is what the failure is. Which errno the console answers is unmeasured, so the test pins the family and the sign - what a caller branches on - not a code no run has established.

What it opened

A/B on one build, ORBISTOUN_RETURN=sceKernelStat:0x7fff0001 forcing the stub back:

with stat   188 distinct, 415,532 calls
as a stub   181 distinct, 415,359 calls

only once stat answered:
  sceKernelOpen, sceKernelClose, sceKernelPread, sceKernelAllocateDirectMemory,
  sceKernelDeleteSema, and three libSceAmpr command-buffer calls

The guest starts opening and reading files and using the decompression library - the asset-loading path, sitting behind one unanswered question about whether a file was there.

One function stopped being called, and that is the corroboration

Eight gained, one lost - sceKernelMkdir - net seven. The guest stops creating a directory because stat now tells it the directory exists. Worth more than the eight gained: a wrong implementation adds calls, but it does not make a guest stop doing something that has become unnecessary.

Two gates caught what I did not

every_implemented_function_is_written_down - the knowledge entry was missing.
every_implementation_is_also_declared_here_or_says_why_not - "sceKernelStat is implemented but not declared". The body belongs in metadata, beside the POSIX form it shares a success path with; the name belongs to libkernel_fs, where it is declared. Registering it beside the POSIX ones offered it under the wrong library - a real structural error that nothing in the change itself would have surfaced.

Break watched to fail

("sceKernelStat", stat)  ->  "NOT the POSIX -1 - that is the whole reason this is not the
                              POSIX function under a second name"        FAILED

Decision: D525.