2026-08-20 (evening) - The wall had a name, and it was C++

D124-D126. Named the biggest wall, implemented the C++ static-initialisation runtime, replaced the hand-typed word list with a cited harvest, and found the systemic stub bug underneath three unrelated faults.

Surprises

• The 53.5% import was __cxa_atexit. Found by reasoning about the shape of the traffic - a libc function called more than all others combined is allocation or static initialisation, both of which are C++ ABI symbols no C word list produces - rather than by searching wider.
• __cxa_guard_acquire was worse unimplemented than absent. An error return is non-zero, which reads as "go ahead and initialise", and with release doing nothing the flag never set. Every function-local static reconstructing on every visit, forever.
• An error code in a pointer register is a wild pointer. Three titles faulting at three unrelated addresses shared one cause: functions that return handles were answering with error codes, which the guest dereferenced. Only visible from the whole call list - the worklist ranks by volume, and the function that ended the run was called twice.
• My own harvest filter dropped the most important name in the corpus. "Skip reserved names" excluded __cxa_atexit. FreeBSD already marks implementation detail with FBSDprivate_* blocks; a rule the source states beats one invented on top of it.
• A citation named a temporary directory. Useless to anyone re-deriving it. The revision leads now and the path is dropped.
• orbistoun-cli could not build for an hour, because it links the whole workspace and another session was mid-edit in the shader translator. The harvest now also exists as an example in a crate that depends on nothing but hashing.

Outstanding

read of 0x5 did not move after marking scePthreadSelf as pointer-returning, so either the hypothesis is wrong or the wiring is not reaching it - unverified either way, and the next thing to check. Thread pointer and previous-generation containers still untouched.