Extract built-in metadata helpers out of pipeline planning

[hynky1999]

Background

PR #214 was merged as a hotfix for the mdr.inference import cycle by deferring the TaskStep import in refiner.pipeline.planning.

That fixes the immediate failure, but the underlying design smell remains: inference and robotics built-ins import describe_builtin from refiner.pipeline.planning, even though describe_builtin is just metadata/decorator plumbing. Importing the planner should not be required to tag a callable as a Refiner built-in.

The failure path that motivated the hotfix was roughly:

mdr.inference
-> refiner.inference.generate_text
-> refiner.pipeline.planning
-> refiner.pipeline.sources.task
-> refiner.pipeline.sources.__init__
-> refiner.pipeline.sources.readers
-> refiner.pipeline.sources.readers.lerobot
-> refiner.robotics.motion
-> refiner.pipeline.planning.describe_builtin

At that point planning.py was partially initialized, so describe_builtin was not available yet.

Proposed direction

Extract built-in metadata helpers out of refiner.pipeline.planning into a dependency-light module, for example:

refiner.pipeline.builtins

That module could own:

• _REFINER_BUILTIN_CALL_ATTR
• describe_builtin(...)
• possibly shared helpers for reading/parsing built-in metadata, if useful

Then update built-in producers/consumers to import from the lightweight module instead of duplicating the attr string or importing the planner:

• refiner.inference.generate_text
• refiner.robotics.motion
• refiner.robotics.hand_tracking
• refiner.inference.internal.runtime
• refiner.services.discovery
• refiner.platform.manifest
• refiner.pipeline.planning

For compatibility, refiner.pipeline.planning can continue re-exporting describe_builtin for one release cycle, but it should no longer be the ownership location.

Acceptance criteria

• Importing mdr.inference in a fresh Python process succeeds.
• Inference and robotics built-ins do not import refiner.pipeline.planning just to access describe_builtin.
• The built-in metadata attribute name is centralized instead of duplicated across modules.
• Existing callers of from refiner.pipeline.planning import describe_builtin either keep working via a documented compatibility re-export or have a deliberate migration path.
• Regression coverage uses a subprocess import test so partial-initialization cycles are not hidden by pytest import order.

Suggested regression:

subprocess.run(
    [sys.executable, "-c", "import refiner as mdr; mdr.inference"],
    check=True,
)

[mvanhorn]

Implemented in #220 - refactor: Extract built-in metadata helpers into a dependency-light module.