fix: convert a character_not_in_repertoire error to an invalid-class error

[grempe]

Closes #854.

@g3nke11 claimed this on 2026-09-17. I linked this branch on the issue on 2026-09-22 and offered to open it if they had moved on; there has been no reply since, so I am opening it now. Happy to close it if they would rather finish their own.

PostgreSQL rejects text that is not valid UTF-8 or contains a NUL byte with 22021 character_not_in_repertoire. Every Postgrex.Error clause in handle_raised_error/4 ends in Ecto.Adapters.Postgres.Connection.to_constraints/2, which finds nothing for that code, so a create, update or filter with such a value returns Ash.Error.Unknown, and the API extensions answer 500. Only the data layer knows this storage has that limit; ETS and Mnesia store the value.

This adds a clause matching %Postgrex.Error{postgres: %{code: :character_not_in_repertoire}}. It matches on the error code, not on message text. It sits before the existing Postgrex.Error clauses for the {:bulk_create, _} and {:ecto_changeset, _, _} contexts, which would otherwise capture it first.

• In a query it becomes Ash.Error.Query.InvalidFilterValue.
• In an update it becomes Ash.Error.Changes.InvalidAttribute for each change whose value fails String.valid?/1 or contains <<0>>. The error names no column, so the attribute comes from the changes.
• In a create it becomes Ash.Error.Changes.InvalidChanges. The rescue there only has the placeholder changeset that bulk_create/3 builds for it, with no changes, so the attribute cannot be named.

The message is Postgres's own (invalid byte sequence for encoding "UTF8": 0x00). One thing I would change once the no_value? option you suggested on #863 exists: the query case has no value to put on InvalidFilterValue, so it currently renders as Invalid filter value `nil`: ..., the same as #863. Happy to adjust it here along with that.

Tests: test/repertoire_error_test.exs covers a filter, an update with a NUL byte and with invalid UTF-8 (asserting the attribute), a create, and an ordinary-text control. On main the four error cases return Ash.Error.Unknown; with the change they pass. Full suite 1011 passed on main, where the same three JoinSubquerySortTest / UniqAggregateSortTest cases from #858 fail with and without this change, and 983 passed with the patch on v2.13.1. mix format --check-formatted, mix credo --strict, mix sobelow and mix dialyzer clean.

The branch is based on current main, and it also merges cleanly with #863 in either order; I checked main plus both together, and the full suite passes on that with only the three #858 failures.

Found by an AI agent working with a human fuzz-testing their own application; the reproduction outside this repo is https://github.com/grempe/ash-fuzz-repros/blob/main/test/ash_postgres/nul_byte_in_text_test.exs, where all eight cases pass against this branch, across Ash, AshJsonApi, AshGraphql and ash_lua.

Contributor checklist

Leave anything that you believe does not apply unchecked.

• I accept the AI Policy, or AI was not used in the creation of this PR.
• Bug fixes include regression tests
• Chores
• Documentation changes
• Features include unit/acceptance tests
• Refactoring
• Update dependencies

[zachdaniel]

🚀 Thank you for your contribution! 🚀