Self-match twin tests assume one pandas missing-string representation #459

The full release gate for PR #456 found two failures in the new direct/window self-match twin tests: self_gene_id uses pandas' missing string sentinel (NaN) in inferred-string mode, but the test expected literal None. The API retained the expected missing value; its asserted representation was too specific. The other 5,044 tests passed.

Assert the missingness mask and the nonmissing origins separately. Run both paired doors explicitly under the shared pandas string-mode fixture, then repeat the full gate. No production behavior or warning suppression changes are needed.