2026-08-21 - A shape, not a word list (D189), and an audit that was already red

The prefix-posix shape named the whole scePthreadAttr* family in one pass - fifteen of sixteen new names, including scePthreadExit, Yield, MutexDestroy and MutexTrylock. One rule, sce + the harvested POSIX name camel-cased, derived rather than written.

I overstated D168 and have corrected it in place. D168 said the harvest was missing a syscall family, and that is true - socket, setsockopt, sched_yield are genuinely absent, because the harvest read only four library directories. It simply was not the reason the mutexattr family was unreachable; every token that family needs was already harvested.

Worth naming as a failure mode: a correct known problem is the most convincing wrong explanation available. It fits the shape of the question, it is already written down, and nothing about it looks like a guess.

The symbol audit fails, and did before this session

orbistoun-cli audit symbols/generated.json reports 218 names it cannot account for. Every one has a derivation record; the records are simply unverifiable - they claim crates/orbistoun-names/data/standard.txt and that file does not contain them. Two groups:

• ~187 syscall wrappers absent because the harvest is incomplete. D168's fix - re-harvest from a fuller checkout - is the answer, and needs a FreeBSD tree.
• 31 C++ ABI names (_Unwind_Resume, _ZTVN10__cxxabiv1..., __cxa_*) which are not FreeBSD at all. They are citable to the Itanium C++ ABI; their record cites the wrong source. A vocabulary gap in the provenance vocabulary rather than the name one.

Not introduced here - the same 218 fail with or without today's merge, which added 16 names that all account for themselves. Recorded because the symbol-provenance CI job runs this and would be red, and a check nobody has looked at is worth less than no check.