D127 - Two invented rules in one afternoon, and the tests caught both

decided · 2026-08-20

D126 recorded one rule I invented on top of a source rather than reading it from the source. There were two, and the second was found only because a test refused to pass.

The harvester looked for files named Symbol.map. FreeBSD's threading library declares its exports in pthread.map, so a harvest that announced 2,497 names successfully had lost every pthread_* symbol - including pthread_create and pthread_mutex_lock, both of which real titles import.

Now any .map file: the format is what makes a file relevant, not what somebody named it. 2,497 became 2,637, and matches against a real import table went 123 to 125.

What actually caught it

A test asserting that four named functions survive the harvest. It was written as a shape check and turned out to be a coverage check - pthread_create failing is what exposed the missing library.

It is now explicitly one name per library, because that is the property worth holding: a harvest that silently drops an entire library still reports success, and the total looks plausible either way. 2,497 is not obviously wrong. Only a missing name is.

And one absence that is correct

clock_gettime is in no version script at all - FreeBSD generates syscall stubs from syscalls.master at build time. The test now asserts it is absent, with a note saying why, so that if it ever appears somebody is told the reasoning is stale rather than left to wonder.

That is the honest shape for a known gap: pin it, explain it, and make its disappearance noisy.

The uncomfortable part

I reported these crates clean while two tests were failing. My verification loop counted lint warnings and summed test totals, and never looked at failures - so a red suite summed to a number and passed inspection.

A check that cannot fail is not a check. Both of the afternoon's real bugs - the reserved name filter and the filename - were found by tests I had already written and then not read the results of.