test: gate the cache-name conflict, and say so where it happens (#706)

[realmarcin]

Closes #706.

What the issue asked, and what I found instead

#706 asked whether the normaliser's discarded exit code should fail a recipe. Working it, the exit code turns out to be the least of it — and correctly left alone.

The committed tree was only gated on Linux. test_reference_cache_names_are_case_exact.py reads os.listdir(REFERENCES_CACHE) — the filesystem. macOS cannot hold DOI_x.md and doi_x.md at once, so on a developer machine the pair is invisible, and the divergence appears only in CI.

That is #690's lesson stated as a test: a green local run is not a green CI run. The new check reads git's index, which records the name it was told regardless of what the filesystem will store, so the clash is caught everywhere.

Mutation-checked by staging a clashing name straight into the index:

git update-index --add --cacheinfo 100644,<blob>,references_cache/PMID_10049867.TXT
→ both tracked; test RED
→ git update-index --force-remove ...; test green

A state macOS cannot represent on disk — which is precisely why the index is the right thing to read.

The exit code, deliberately unchanged

The recipes' exit $code is the validator's, and that is right: normalising must never turn a failing validation green (#697). So the exit code cannot be the signal. Instead the script prints a trailing banner naming the count and what a human has to decide, so a conflict is not one line lost in a long validation log.

Tested by stubbing rename, because the real conflict needs two files whose names differ only by case — the filesystem, not the code, is what makes it unreachable here. Both directions asserted: a conflict prints it, no conflict does not, and removing the banner reds the first.

just lint       clean
pytest tests/   2912 passed, 23 skipped

Known unrelated failure: test_no_snippet_stops_mid_word, from an untracked local cache on a record this PR does not touch.

Not done

Making just qc carry the conflict, one of the issue's suggested alternatives. It is unnecessary now: the committed tree is gated by a test that runs in validate-strict, which is a blocking lane, and qc-references is explicitly not a CI gate.