2026-09-01 - obSCEne is the oracle; flexible-memory fixed byte-exact against hardware
User's steer: obSCEne passes on hardware, so its crash under orbistoun is an orbistoun bug - and it's the ideal oracle (555 checks with recorded hardware answers). Two parts.
Stale eboot: titles/obscene/eboot.bin was yesterday's
build and crashed early - its backward ELF-magic base scan ran off the
mapping because that build didn't import
sceKernelVirtualQuery, disabling its own guard (not chiefly
orbistoun's fault; a vendor module's header is legitimately not resident
at the base). selfish builds again (the user deployed obSCEne today), so
rebuilt make module fresh and installed it as the title. It
now runs 338 checks under orbistoun.
Flexible memory, measured exactly by the oracle vs
console-klog.01092026.txt: flexible-available
was answering ~0x1_3f01_0000 (the direct pool) where
hardware says 0x1b40_0000, and
flexible-configured failed (unimplemented) where hardware
says 0x1c00_0000. Fixed: flexible is now a separate budget
in direct (configured 0x1c00_0000, launch
available 0x1b40_0000, atomic mapped counter);
sceKernelConfiguredFlexibleMemorySize implemented;
available reads the budget minus mapped;
map/release charge/credit it and
map refuses past it. Both are the system default (no title
overrides the mem-param, D442) so they transfer. Re-ran obSCEne under
orbistoun: both checks now match hardware byte-exact
(available 0x1b400000, configured 0x1c000000
pass). Unit test pins the budget arithmetic; kernel/mem tests pass,
clippy clean, fmt-clean (D444).
Next wall: obSCEne faults at image+0x4502ff reading
0x600000801000, one page past the guest stack - a later
probe walks off the stack. That's the next investigation.