Look up namespaced SQL tables by their schema-qualified key

[DMZ22]

• fixes KeyError: 'table_name' when writing a resource to postgres table in schema #1602

---

Writing a resource to a table inside a namespace (schema) raised KeyError:

resource.write(sql_url, control=formats.SqlControl(table="reprex", namespace="datapackage"))
# KeyError: 'reprex'   (frictionless/formats/sql/adapter.py, write_row_stream)

Cause. SqlAdapter.__init__ builds MetaData(schema=self.control.namespace). When a schema is set, SQLAlchemy keys every table by its schema-qualified name — Table.key is "<namespace>.<name>", and metadata.tables is keyed the same way:

>>> md = sa.MetaData(schema="ns")
>>> t = sa.Table("reprex", md, sa.Column("x", sa.Integer))
>>> list(md.tables)        # ['ns.reprex']
>>> t.key                  # 'ns.reprex'
>>> md.tables["reprex"]    # KeyError

But every lookup in the adapter used the bare name — self.metadata.tables[table_name] — so write_row_stream, read_schema, read_cell_stream, delete_resource and write_resource_with_metadata all KeyError the moment a namespace is in play.

Fix. Route the six lookups through a small helper that prepends the namespace when one is configured, matching how the tables are actually keyed:

def _table_key(self, table_name: str) -> str:
    if self.control.namespace:
        return f"{self.control.namespace}.{table_name}"
    return table_name

No behaviour changes when namespace is None (the helper returns the name unchanged), so the existing non-namespaced paths are untouched.

Verification

Reproduced and fixed end-to-end on SQLite with an attached schema (no server needed):

without fix:  KeyError: 'table'   at adapter.py write_row_stream
with fix:     wrote + read back -> [['id','name'], [1,'english'], [2,'中国人']]

Added test_sql_adapter_namespace_issue_1602 in the adapter spec's Bugs section — it attaches a second SQLite database as a schema, writes a package into it, and reads the rows back. Reverting only adapter.py while keeping the test makes it fail with the original KeyError, so it guards the fix. The full SQL adapter suite is 38 passed / 0 failed (59 skipped are the server-backed backends). ruff format --check is clean, and ruff check reports the same count on adapter.py before and after, so this adds no new lint.