Docs follow-ups from the #777 panel: run the Postgres-only blocks (F25), issue numbers in error messages, refused models stay registered #781

Problem

Three things surfaced while fixing the review panel's findings on the docs-rewrite PR (#777) that are outside that fix round's scope.

Items

1. Postgres-only blocks still compile-only (panel F25). About a dozen page blocks (concepts/backends.md, guide/connections.md, guide/raw-sql.md, guide/row-level-security.md, guide/schema/auto-migrate.md, and Session.set_config's docstring) are marked {test=compile reason="needs a PostgreSQL server"} although the suite has a live Postgres under the postgres_only marker. Running them needs a fence setting such as {test=run prelude=<name> backend=postgres}, a Postgres twin of the preludes that connect (a patched connect() URL), the fence tests split so the Postgres half runs in the backend-matrix job, and Litestar/FastAPI installed for the two framework blocks. Story 4 of Docs rewrite: every snippet runs, every fact matches the code, and the API reference renders cleanly #698 asks that every copyable block be executed.
2. Error messages carry issue numbers into the docs. The printed-output gate (docs: panel round 1 on #777 (harness and pages) — output blocks regenerated, one fence reader and runner, twin equivalence, CLI reference complete #780) regenerates error blocks from Ferro's real messages, and several contain references such as (#291). The pages rule (and I-14 for docstrings) keeps issue numbers out of reader-facing text, but these come from src/. Decide whether exception texts should name issues at all; if not, remove them from the messages (the regenerated blocks then update themselves).
3. A model Ferro refuses at class definition stays in the registry. Found while regenerating the aggregation error blocks: when the metaclass raises for a bad declaration, the partially built class remains registered, so a later create_tables() or lookup can meet it. Looks like a src/ferro/registry.py / metaclass bug; a refused class should leave no trace.

Related

• Docs rewrite: every snippet runs, every fact matches the code, and the API reference renders cleanly #698 (docs rewrite), PR docs: the docs rewrite — every page one kind, every block runs, the SQL beside every query (#698) #777 panel report, PR docs: panel round 1 on #777 (harness and pages) — output blocks regenerated, one fence reader and runner, twin equivalence, CLI reference complete #780 (where 1 was declared not done and 2–3 were observed).