D191 - The harvest walker ignored the rule written to fix it

decided · 2026-08-22

Re-harvesting from a FreeBSD checkout, to close the provenance gap D168 describes, lost 147 names and added six. A harvest that reports success while shrinking is the shape worth chasing, and it led to a defect older than the exercise.

Two implementations of one rule, and only one was fixed

orbistoun_names::harvest::is_version_script accepts any .map file. Its doc comment records why: libthr calls its exports pthread.map, and a rule that looked for Symbol.map cost every pthread_* name (D127).

The harvest command does not call it. It tests file_name() == "Symbol.map" directly, and has since before D127 was written. The fix landed in the crate and in an example binary; the command that anybody actually runs kept the original rule and kept reporting success.

So the walker skipped lib/libthr/pthread.map, lib/libsys/Symbol.sys.map, Symbol.thr.map, syscalls.map and every per-architecture map - ten files of fifty-seven. Wired to the rule that already existed: 2685 names to 3064, and the two "lost" entries are words from the old file's own header.

lib/libsys also had to be added to the directory list: FreeBSD moved its system-call stubs out of libc, so socket and its neighbours left from under a constant naming lib/libc.

D168 was right about the symptom and wrong about the cause, and so was I

D168 said the harvest was missing a syscall family. True. Its proposed fix - re-harvest from a fuller checkout - was not the fix: depth and completeness were never the problem. A directory had been renamed underneath one constant and a filename convention had never been honoured by another.

Two predictions made before doing the work, both wrong, both by generalising from the head of a list. Reading five names off the top of the audit's output, I said it would turn "~187 records" green: it turned a handful green, because the unaccounted are 202 vendor sce* names and 17 C++ ABI names, not syscall wrappers. And I said the enlarged vocabulary would plausibly name several of the corpus's unnamed imports: it names none of the twenty-six.

The work was still worth doing, for a reason neither prediction named - it surfaced a defect that had been silently degrading every harvest since D127.