2026-09-01 - Fault reports carry the faulting instruction (D432)
(worker) The fault report now prints the bytes AT the instruction
pointer and a 48-byte window BEFORE it, read straight from the loaded
image inside the handler - page-clamped both ways so the read cannot
fault, allocation-free, one write (D065: executable pages are readable
here). Line capacity 1024 so the register/frame lines are
never the part dropped. This is what a
dump/disasm verb would have been, but
automatic: a SELF/eboot fault, whose image+0x… the ELF
offsets cannot locate, is now a disassembly on the page.
Used it on the shared PPSA02664/PPSA25872 wall and it disassembled
cleanly: the guest calls image+0x7b5890, that
returns 0, and the guest dereferences it unchecked -
mov [rax],rax with rax=0, initialising a self-referential
(empty circular-list) structure through a null pointer. The three
lea [rip+…] right before the fault compute exactly the
dump's rsi/rdi/rdx, confirming the decode. So the "just before" memory
calls were a red herring; the wall is a guest function returning null.
Honestly ruled out ORBISTOUN_MAP_SHAPE=reserved-low
(non-zero physical base) - it changed nothing.
The wall is now the precise question "why does image+0x7b5890 return
0" rather than an address, with the tool to answer it in place. Not
drilled further: naming that function or blind-stubbing
_init_env (no citable contract) would be the
pointer-vs-error guess the finding warns against. Worker tests green
(40+9+1); kernel green (74+33+33); my code clippy-clean (orbistoun-fs
escape/socket debt is pre-existing).