What

Brings the use_kernel=True path closer to Thrift parity on the cursor-state-tracking (T7) and metadata clusters from the API+CUJ gap audit, surfaces DML rowcount, and makes async command state server-sourced (no connector-side state to drift). Batch 3 of the use_kernel parity work (follows #824 / #830).

Depends on kernel PR databricks/databricks-sql-kernel#139 (case-insensitive table_types, HTTP error-envelope parsing, num_modified_rows getter, Session::attach_async_statement). KERNEL_REV points at that PR's branch HEAD; re-pin to the squash SHA before this merges.

Async command state — server is the source of truth

This addresses the core concern that connector-side async bookkeeping drifts from the server.

• get_query_state / get_execution_result re-attach by id. They call Session.attach_async_statement(command_id) and read state / await results from the server, instead of consulting a connector-side handle map. SEA keys GetStatementStatus/Result purely on the id, so this works whether or not the connector still holds the submitting handle — and get_execution_result is inherently re-callable (each call re-attaches), matching Thrift's re-fetchable operation handle. No more unknown command_id on a second fetch.
• Deleted _closed_commands (and its bounded-FIFO machinery _record_closed / _CLOSED_COMMANDS_MAX). This was the invented state — the connector guessing CLOSED-vs-SUCCEEDED locally, with the bounded FIFO making the answer depend on how many other commands had closed since. The server reports CLOSED natively (verified live: GET after close → 200 state=CLOSED until the result TTL, then 404 NOT_FOUND → mapped to a terminal state via _is_not_found).
• _async_handles / _async_statements demoted to keep-alive-only. They're retained for one reason: the submitting ExecutedAsyncStatement's Drop fires close_statement, which would kill the live async query — so the connector holds the handle until close_command / close_session. They are no longer consulted for state or results.

Cursor-state tracking (T7)

• Metadata no longer leaves active_command_id stale. Set in the _make_result_set chokepoint so sync execute, async fetch, AND metadata all leave the cursor pointing at the producing command (matches Thrift's _handle_execute_response).
• Failed sync execute publishes the statement id (from the canceller's inflight slot) before re-raising. Best-effort; never masks the original error.

Metadata

• Wildcard/blank catalog → all-catalogs ('%'/'*'/''/None → None) across schemas/tables/columns (_catalog_or_none), making them symmetric (fixes columns).
• Empty-string pattern → match-all (_none_if_blank); %/* stay as LIKE wildcards.
• Dropped the table_types connector drain — the kernel filters it (case-insensitively, Bump version to 2.6.0 #139). Removed dead _drain_kernel_handle/_StaticArrowHandle + unused imports.
• Fixed Cursor.columns() docstring — catalog_name=None is accepted on all backends.

DML rowcount (enhancement)

• Surface num_modified_rows as cursor.rowcount; None leaves -1 (matches Thrift). getattr-guarded.

Testing

• Unit (tests/unit/test_kernel_client.py, 116 pass): get_query_state/get_execution_result rewritten to mock attach_async_statement (server-state, NotFound→terminal, re-callable); close paths assert close+drop (no _closed_commands); metadata normalization; num_modified_rows→rowcount. test_client.py 128 pass.
• E2E (live dogfood): async execute_async→poll→fetch, get_async_execution_result re-callable, result resumable from a fresh cursor (proves state comes from the server, not per-cursor tracking); empty-string filter; case-insensitive table_types (self-contained table+view); active_command_id after metadata; DML rowcount. All green.
• black (pinned 22.3.0) clean.