39 cache files use prefixes no citation resolves, and it is unclear whether anything reads them

[realmarcin]

Summary

references_cache/ holds 39 files whose prefix corresponds to no reference in
the corpus: 20 pmc_, 9 europepmc_, 6 epmc_, 3 openalex_, 1
semanticscholar_. Nothing in this repository writes them and nothing resolves
them — every reader builds a stem from a doi: or PMID: citation, and those
are the only two prefixes any record uses.

Recorded rather than acted on, because "unused" and "safe to delete" are not the
same claim and I could not establish the second.

Evidence

$ grep -rhoE 'reference: *"?([A-Za-z]+):' kb/ data/ | sort | uniq -c
   4716 PMID:
   1096 doi:
      2 bioproject:

$ git ls-files references_cache/ | sed 's|.*/||' | grep -oE '^[A-Za-z]+_' | sort | uniq -c
    497 PMID_
    212 doi_
     20 pmc_
      9 europepmc_
      6 epmc_
      3 openalex_
      1 semanticscholar_

No script constructs or reads an uppercase PMC_-style name, and
grep -rn "pmc_\|europepmc_\|openalex_\|semanticscholar_" scripts/ src/ returns
only unrelated function names (openalex_locations, apply_pmc_conversions).

Why it came up

PR #702 adds scripts/normalize_cache_names.py, which renames cache files whose
prefix casing no citation resolves to. My first version listed all seven prefixes
as canonical — asserting a correct casing for five that have no citation to
derive one from, and putting 39 files inside the renamer's blast radius on the
strength of a guess. They are now in UNRESOLVED_CACHE_PREFIXES: neither renamed
nor flagged.

The questions worth answering

1. Where did they come from? Probably an earlier fetching path, or a sibling
   repo's convention. git log --diff-filter=A -- references_cache/pmc_* should
   say.
2. Does anything outside this repository read them? The cross-Mech tooling
   (culturebotai-claw, CultureMech) shares conventions with this cache; a file
   unused here may not be unused there.
3. If the answer to 2 is no, should they be deleted? 39 files of committed
   evidence is not much weight, but a cache that nothing reads is a cache nobody
   is maintaining, and it dilutes references_cache/ as a place where every file
   means something.

Not urgent

They cost nothing but space and a little confusion. The reason to resolve it is
that the next person writing a cache-wide tool has to make the same judgement
call I did, with the same lack of information.

Acceptance test

Either the five prefixes are documented with what reads them, or the files are
removed and UNRESOLVED_CACHE_PREFIXES shrinks to empty.