CONTRACT 5: the visibility attribute's `visibility_type` carried the dtype word
("bool"), where the schema's `visibility_types` allows only "boolean" or "ratio".
The migrator now rewrites it from the pairing the schema itself declares, the writer
emits that pairing, `check_attributes` asserts the enum and the dtype/word pairing
(it previously proved only presence and type), and `verify.compare_metainfo` accepts
that one rewrite and nothing else. All 87 metainfo files (174 attributes) were
regenerated by the migration itself, not hand-edited. No media was re-encoded: the
12 review pages change by exactly those two enum values (+6 B each) and their
embedded images are byte-identical.
`validate.py`: a `--views`/`--limit` run no longer asserts the release-wide census
(87 views, 41,258 train / 2,728 test frames). A subset run validates its selection,
reports `coverage: "partial subset"`, writes `tmp/rhd-tools/validate-partial.json`
instead of the acceptance report, and leaves the releas…
