Summary

Makes WorkspaceClient.dbutils a functools.cached_property so consumers that never read it pay no construction cost — and, on Spark Connect runtimes, never touch the legacy SparkContext path that databricks.sdk.runtime materializes on import. Includes four regression tests that lock in the contract.

Why

WorkspaceClient.__init__ used to call _make_dbutils(self._config) eagerly, which on a cluster imports databricks.sdk.runtime. On a Spark Connect (shared-access-mode) cluster, that import materializes a legacy SparkContext and raises CONTEXT_UNAVAILABLE_FOR_REMOTE_CLIENT, crashing the constructor before any API call. Downstream consumers that never touch .dbutils (notably dbt-databricks Python models) hit this for no reason — see #1463 and databricks/dbt-databricks#1252.

#1469 patches the runtime side as a defense-in-depth fallback (catch the materialization failure, fall back to RemoteDbUtils). This PR is the durable fix: callers that don't read .dbutils never trigger the build at all, sidestepping the entire code path. The first read still calls _make_dbutils once, lazily; subsequent reads hit the cached attribute in __dict__ at plain-attribute speed.

What changed

databricks/sdk/__init__.py (generated from updated template):

• from functools import cached_property added to the imports.
• The eager self._dbutils = _make_dbutils(self._config) line is removed from __init__.
• @property def dbutils (which returned the cached self._dbutils) becomes @cached_property def dbutils that calls _make_dbutils(self._config) on first access.

_dbutils was a private attribute with no external consumers (verified across the codebase), so removing it does not break any public surface.

tests/test_client.py — four new tests:

• test_dbutils_is_a_cached_property — descriptor type check.
• test_workspace_client_init_does_not_build_dbutils — spies _make_dbutils, constructs a WorkspaceClient, asserts the spy was never called.
• test_dbutils_first_access_builds_exactly_once — first read invokes _make_dbutils once (returns the spy's sentinel); second read still shows call_count == 1 and same identity.
• test_workspace_client_constructs_on_spark_connect_without_touching_runtime — fakes dbruntime to raise CONTEXT_UNAVAILABLE_FOR_REMOTE_CLIENT on any namespace materialization; asserts WorkspaceClient(config=...) succeeds and databricks.sdk.runtime is never imported during construction. This is the strongest evidence that the dbt-databricks failure mode is sidestepped by this change alone.

How is this tested?

• 4/4 new tests pass locally (0.03s).
• Existing tests/test_client.py autospec tests untouched, still pass.
• The fourth test is the negative-space proof: asserts databricks.sdk.runtime is not in sys.modules after WorkspaceClient(config=...) — i.e., the constructor literally does not reach for the runtime module.

NO_CHANGELOG=true