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, and surfaces DML rowcount as an enhancement. This is 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 pyo3 getter). KERNEL_REV here points at that PR's branch HEAD; it will be re-pinned to the squash SHA before this PR merges.

Cursor-state tracking (T7)

• Metadata no longer leaves active_command_id stale. Set it in the single _make_result_set chokepoint so every result-producing path — sync execute, async fetch, AND metadata — leaves the cursor pointing at the command that produced the current result set (matches Thrift's unconditional set in _handle_execute_response).
• Failed sync execute publishes the statement id. Read the server-issued id from the canceller's inflight slot and set active_command_id before re-raising, so a failed query is still correlatable (telemetry / log lookup). Best-effort; never masks the original error; untouched on a pre-id transport failure.
• get_async_execution_result() is re-callable again. Removed the eager close/drop of the async handle after the first await_result() — the kernel's await_result() is idempotent, so the handle stays tracked until close_command/close_session (Thrift parity). The prior eager close made a second call raise.
• _closed_commands is bounded. Now an OrderedDict FIFO capped at 10,000 (_record_closed) instead of an unbounded set.

Metadata

• Wildcard/blank catalog → all-catalogs. '%' / '*' / '' / None normalize to None across get_schemas/get_tables/get_columns (_catalog_or_none), making the three symmetric (fixes columns, which treated catalog as an exact identifier). JDBC exact-or-all semantics.
• Empty-string filter → match-all. Whitespace-only pattern args normalize to None (_none_if_blank), mapping the kernel's InvalidArgument-on-"" to Thrift's effective match-all. %/* stay as real LIKE wildcards on patterns.
• Dropped the table_types connector drain. The kernel now filters table_types itself (case-insensitively, kernel Bump version to 2.6.0 #139), so get_tables forwards it and returns the stream unchanged — restores streaming and removes the duplicated case-sensitive client-side filter. Deleted the now-dead _drain_kernel_handle / _StaticArrowHandle helpers + unused imports.
• Fixed the Cursor.columns() docstring — catalog_name=None is accepted on all backends (not rejected with ProgrammingError).

DML rowcount (enhancement)

• Surface the kernel's num_modified_rows as cursor.rowcount for DML instead of the hardcoded -1. None (SELECT, or warehouses that don't report it) leaves it at -1, matching Thrift. NB this makes the kernel path exceed Thrift, which also hardcodes -1. getattr-guarded against an older wheel.

Testing

• Unit: tests/unit/test_kernel_client.py — added/updated tests for A–D (active_command_id chokepoint + failure-path + bounded _closed_commands + re-callable get_execution_result), metadata normalization (wildcard catalog, blank pattern, kernel-side table_types forwarding), and num_modified_rows→rowcount. 115 pass.
• E2E (tests/e2e/test_kernel_backend.py, live dogfood): empty-string filter match-all; case-insensitive table_types (self-contained create-table+view, lowercase view / uppercase TABLE); active_command_id published after a metadata call; DML rowcount wiring doesn't break DML. 4 pass in ~20s.
• black (pinned 22.3.0) clean.