test(compilers/openapi): stop the reproducer tables drifting

[OmarAlJarrah]

Summary

danglingDir's doc comment said the directory holds twelve issue-#14
reproducers. It holds fourteen, and the test's own table names all fourteen:

$ ls testdata/dangling/openapi/*.yaml | wc -l
14

f31-discriminator-empty-name.yaml and f32-ref-noncanonical-escape.yaml are
the two the count predates. The number was true when written and rotted as
fixtures were added.

The count is dropped: a count in a comment beside the list it counts has no way
to stay true, and the table below it is the enumeration.

Dropping the number leaves the drift it came from, though — the table itself is
hand-maintained, so a fixture copied into the directory and not added to the
table is compiled by nothing and the suite stays green while the corpus grows
past it. So the table moves into danglingReproducers() and a new test holds it
to the directory in both directions. That is what the comment now points at
instead of a number.

Test plan

TestDanglingRefs_EveryReproducerIsExercised, confirmed to redden against a
planted defect in each direction rather than accepted because it went green:

• Fixture on disk, absent from the table — deleting the f32 row fails with
  f32-ref-noncanonical-escape.yaml is in ../../testdata/dangling/openapi but not in danglingReproducers.
• Table naming a fixture the directory lacks — adding an f99-ghost.yaml row
  fails with should have 14 item(s), but has 15. That length check catches both
  directions, so its message names neither: it fires beside the check above when
  the directory is the side holding the extra, and saying "the table names a
  fixture the directory does not hold" there would be false.

The same drift, at its other two sites

cycleReproducers is the same shape: a hand-maintained list of cycle_*.yaml
fixtures, kept twice — in compilers/openapi for the compiler's refusal, and in
internal/scan for the pre-parse scan's diagnosis — with nothing holding either
to the directory. Both were complete when checked (24 rows against 24 fixtures,
verified in both directions), so nothing had drifted yet; what was missing was
anything to stop it.

Both now carry the twin of the guard above. Holding each to the same directory is
also what keeps the two tables from drifting apart, which neither package can
check directly since a test package cannot import another's. Each package spells
its fixture path once as reproducerDir, so a guard and its reader cannot
disagree about where they are looking.

Planting a cycle_*.yaml that no table names reddens both packages; adding a row
naming no fixture reddens both on the count.

main is merged in, and the full gate is green on the merged branch: gofmt,
go vet ./..., golangci-lint run (0 issues), go build ./...,
./scripts/check-coverage.sh.

Closes #314