D058 - A check that calls a second library must test that symbol's address itself

Status: bug, found on the host build, generalised.

The harness skips a check whose declared symbol resolved to null, because jumping to zero takes the process down and loses everything behind it. That guard covers the one symbol in the check's table row, and nothing else the body reaches for.

017-posix/spellings-agree calls two libraries by design. On the host build the vendor spelling has no stub, so scePthreadRwlockInit was null, and the check walked straight into it. Segmentation fault, process gone.

Two things worked exactly as intended and are worth naming:

The try record with no res identified it precisely. Not "somewhere in this run" - the exact check. That invariant is the most important property in the program and this is the third time it has paid for itself.

It happened on the host. A null vendor symbol is the normal state on an ordinary machine and an unusual one under an emulator, so the host build met this first - which is the entire argument for having it (D001). Under an emulator it would have appeared as an intermittent crash in a section that usually works.

Fixed by testing every symbol the body calls, since all platform declarations are weak and the address is therefore the test. The general rule: a check that reaches outside its own table row is responsible for the addresses it reaches for. Adding this to the five-step list in CLAUDE.md rather than trusting anyone to remember it.