mutate: refuse oans.c and run_dedupe.c instead of reporting 0% killed on them

[martinus]

Vendored sync of martinus/unordered_dense#205, where the change was made and where the hermetic suite covers it. This is the repository the bug was found in, and the only one of the three that can hit it.

The problem

src/tests.c #includes most of src/*.c into a single translation unit, but deliberately not oans.c or run_dedupe.c — those belong to the shipped binary alone, and make test-build compiles only tests.c.

So a mutant in either file never reaches ./test, and the tool scored it survived:

$ scripts/mutate/mutate.py --file src/oans.c --lines 2172

nothing noticed these - whatever covers them is decoration:
  src/oans.c:2172  == -> !=
      if (ret == ENOMEM || debug)

A dry-run over the whole file planned 1,394 mutants, and every one of them would have come back the same way — a 0% kill rate over 2,177 lines, reported as a fact about the tests rather than about the invocation. That is the flattering direction, and CLAUDE.md's standing rule for this tool is that every way it has been wrong so far flattered the tests.

After

checking the target reaches the binary under test

src/oans.c is not compiled into test, so every mutant in it would come back
`survived` however good the tests are - the build succeeded with an #error in
the file. Sweep a file the test binary is actually built from, or add this one
to it.

The message names the consequence rather than the condition, so the reader is not left to work out why "not compiled in" matters.

How it works

An #error is appended to the target and the build is run. If it still succeeds, nothing the test binary is made of ever included the file. Backend-agnostic on purpose: a compilation database can say which files are compiled, but not which are #included into something that is — and this repository's case is exactly the latter.

Verified here, end to end

• src/oans.c and src/run_dedupe.c — refused.
• src/util.c — still sweeps, still returns real verdicts (delete: clock_gettime(...) survived, as it did before).
• make mutation-replay — unchanged at 65 across eight files, nothing surviving.
• make lint — lint-mutate-core: vendored at c71e0f7a106d, adapter fits.

This is also the make backend, which is one of the two the upstream PR verified end to end (the other being cmake via nanobench; the meson leg is covered by the hermetic suite only, as meson is not installed on this machine).

Note

make lint will fail until martinus/unordered_dense#205 lands, if it is checking the hash against upstream — the mutate_core.sha256 here is updated to match the new copy, which is the convention CLAUDE.md describes (change it there, run that suite, re-copy into all three, update each hash).

🤖 Generated with Claude Code

https://claude.ai/code/session_01GeJoYWWvf2ewg87ih8DpGJ

---

Generated by Claude Code