D416 - Four HLE fixes from the hardware-vs-orbistoun obSCEne diff

measured - 2026-08-31

The latest obSCEne hardware run (a module build, firmware 12.40) diffed against the same suite in orbistoun gave a concrete, ground-truthed defect list. Four fixes, each verified by the verdict flipping when obSCEne re-runs in orbistoun:

• sceKernelWrite(_, _, 0) returns 0 (orbistoun-fs). guest_slice rejects a zero length the way it rejects a null pointer, so a zero-length write returned InvalidArgument - a non-zero value 000-boot/write-returns-count reads as a claim to have written bytes. Handled before the slice, still consulting the descriptor so an unwritable one refuses.
• rand/srand implemented (orbistoun-libc). Unimplemented, rand fell to the placeholder stub and answered the same number every call - 035-libc/rand-seeded catches that as two identical draws. An atomic LCG, seedable so a re-seed reproduces a series; faithful sequence is not what the check measures.
• sceKernelGetSystemSwVersion implemented (orbistoun-kernel). Unimplemented it refused; hardware fills a {size; char[0x1c]; u32 version} struct. Written with the 12.40 this machine models (0x1240_0000), 130-layout/system-software-version now passes.
• sceKernelLoadStartModule returns real handles (orbistoun-kernel). It returned -1 for everything; hardware answers 0x2001 for libkernel, a handle for an /app0 module, and 0x8002_0002 for a /system one. Classified by path so 110-modules/load matches (verdict 0x0 -> 0x3). The honest-failure catch: a first cut handed a handle to any non-system path, so 060-module/load-rejects-missing (which loads a bogus path) reported a nonexistent module as loaded. Corrected to recognise only libkernel, /system, and /app0, refusing anything else.

Two divergences the diff surfaced were deliberately not changed. sceKernelGetModuleInfo still refuses - hardware refuses it too in this run (110-modules/names fail 0x8002_0016), so the principled refusal is correct, not a gap. And the TSC frequency differs by 79 Hz (0x5f259b8e vs 0x5f259bdd), which is boot-calibration variance, not a bug - the lesson there is that obSCEne's 139-exports exact-match on it is too strict, an obSCEne fix.

Recorded measured: each is grounded in a hardware reading and confirmed by re-running the probe.