I think I want to refactor cat_to_filename() to take a buffer and to handle the selective prepending of sys/, perhaps bumping the uexterr_gettext ABI. It's fine for in-tree code, but definitely wrong for out of tree modules.

Having just hit submit... I wonder if we want to spend the available bit in EXTERR_CAT_SRC_MASK on in-tree vs out of tree or maybe just make pre-pending sys/ a legacy category source thing.

This adds two syscalls to each dynamic syscall failure path. I understand that calling the sysctl is the easiest way to handle dynamically loaded modules. But I already got some negative comments due to exterrors adding some nanoseconds to the failing syscalls timings.
Although I do not object in principle, can we have some smarter way to avoid the sysctl if we already queried the specific module? E.g. we can make each module to register 'range' for dynamic errors' pathes, and memoize the range with the values in libc.
This is esp. sorrow because the display of the source line is really a niche feature.

I don't think performance of this path matters at all because users have to go out of their way to enable this code path with EXTERROR_VERBOSE.

That being said, we could easily eliminate one system call by declaring mib _Thread_local and only initializing it once per thread and I suppose one could go further and make it once per-exec.

Simple caching wouldn't be hard, just malloc an kern.exterr.ncategories length array of pointers and use atomics to access/update the slots.

The most optimal implementation would be a no-syscall version using auxargs and shared pages like I mentioned in D58237, but it feels like pre-mature optimization and maximum correctness still requires a fallback like this.