Require permissions on assistant endpoints and validate sampled table names

[chrisclark]

Summary

Fixes a reported security issue in the AI assistant endpoints.

• Missing auth: AssistantHelpView (/assistant/) and AssistantHistoryApiView (/assistant/history/) were plain Views with no permission check, unlike every other Explorer view. The routes are registered even when no EXPLORER_AI_API_KEY is configured. Both now use PermissionRequiredMixin with change_permission, matching the playground, which is where the assistant is used.
• SQL injection: table names from the request's selected_tables were put into SELECT * FROM {table_name} ... via an f-string. table_schema() returned None for unknown tables, but nothing checked that result before sampling rows.
  • build_prompt now resolves each requested table through a new find_table() helper against the connection's schema (which honors the include/exclude prefixes). It skips tables that aren't found and uses the introspected table name.
  • sample_rows_from_table also quotes the name with connection.ops.quote_name as defense in depth.

Tests

• New: anonymous and non-staff users are denied on both endpoints, and run_assistant is never called for them.
• New: build_prompt skips table names that aren't in the schema.
• New: sample_rows_from_table quotes the identifier.
• The new tests fail on master and pass with this change.
• Updated two existing build_prompt tests to patch find_table, and fixed mocks that assigned MagicMock (the class) instead of MagicMock() (an instance).
• The full suite passes under explorer.tests.settings. test_assistant also passes under settings_base.

🤖 Generated with Claude Code