test(corpus): floor the PDF smoke suite's collected pair count

[Nitjsefnie]

Related issue

Closes #601

What does this change?

Adds the completeness floor #598 established for _DIVISION_VERSIONS, applied to
test_pdf_corpus_smoke.py's pair list.

Two assertions, pinning different failure causes:

• len(_PAIRS) >= 15 catches the committed corpus losing PDF fixtures, or adjacent_pdf_pairs()
  ceasing to pair them. 15 sits under the live count of 23 so ordinary fixture churn does not force
  an edit, matching the precedent's ratio.
• len(_PAIRS) == len(adjacent_pdf_pairs()) catches a slice, filter or truncation at the _PAIRS
  assignment itself, with the corpus intact.

Both read _PAIRS rather than calling adjacent_pdf_pairs() afresh. That is the load-bearing
detail: a fresh call guards the corpus, not the list the suite actually parametrizes over. My first
attempt asserted on a fresh call, and _PAIRS = adjacent_pdf_pairs()[:3] — the example from the
issue — then dropped the suite from 138 cases to 18 with the floor still green.

The floor lives in tests/test_corpus_manifest.py rather than beside the suite it guards because
tests/test_pdf_corpus_smoke.py sets pytestmark = pytest.mark.slow at module level and pytest has
no per-test un-marking, so an in-module floor would only ever run in the slow tier. That is the
placement and the reasoning recorded for the committed-fixture floor moved per your #328 review.
The cross-module import follows the existing idiom (tests/test_xml_subsection_nodes.py:52,
tests/test_assignment_classification_boundary.py:874); here it is load-bearing rather than a DRY
convenience, since decoupling it is precisely what made the first attempt guard nothing.

How to test

Red-green, both directions:

• _PAIRS = adjacent_pdf_pairs()[:3] → AssertionError: 3 >= 15 — the floor half fires.
• _PAIRS = adjacent_pdf_pairs()[:16] → AssertionError: 16 == 23 — above the floor, so only the
  equality half fires. This is the case that shows the second assertion is not decorative.
• Healthy tree → passes, collected in the fast tier under -m "not slow and not browser", with the
  smoke module's own 138 cases still deselected. Runs in 0.05s and adds no dependency
  (pypdfium2 is already a core dep).

Known gap, stated because it bounds what this buys: a coordinated slice applied to both the
@parametrize decorator and its ids= list would still escape the floor. The single-site version
of that edit errors loudly at collection, and the same gap exists in #598's floor — I have not
tried to close it here.

Checklist

• Linked the issue above (Closes #...)
• Ran the CI gates locally and they pass (see What CI checks) — run on a runner rather than locally; targeted tests were run locally
• New or changed behavior has tests
• For a bug fix: the test fails without the fix, and I ran it both ways to check
• Disclosed AI assistance below

AI assistance

Generated by Claude Opus 5 (brief, implementation, review, testing)