fix(sampling): migrate budget forcing to mot.generation.usage; drop _meta["usage"] writes

[ajbozarth]

Pull Request

Issue

Fixes #1218

Description

• Migrates think_budget_forcing to read mot.generation.usage exclusively. Raises a named RuntimeError when usage is None, replacing the previous opaque TypeError from arithmetic on None (Ollama's eval_count can be None).
• Drops _meta["usage"] writes from all five backends' generate_from_raw paths now that mot.generation.usage is the unified field. The additive half (writing the field from raw paths) landed with refactor(backends): move generate_from_raw hook firing into Backend base class #1264.
• Replaces an implicit IndexError in the no-sub-calls edge case (think_max_tokens=0 + answer_suffix=None) with a named RuntimeError.

Testing

• Tests added to the respective file if code was changed
• New code has 100% coverage if code was added
• Ensure existing tests and github automation passes (a maintainer will kick off the github automation when the rest of the PR is populated)

Attribution

• AI coding assistants used

---

Adding a new component, requirement, sampling strategy, or tool?

If your PR adds or modifies one of the types below, check the matching box. A checklist of type-specific review items will be posted as a comment.

• Component
• Requirement
• Sampling Strategy
• Tool

NOTE: Please ensure you have an issue that has been acknowledged by a core contributor and routed you to open a pull request against this repository. Otherwise, please open an issue before continuing with this pull request.

[akihikokuroda]

Comment by Claude:

Optional but Recommended: Add Error Case Tests

The PR introduces two new RuntimeError paths that should ideally have test coverage:

1. Missing token counts — When generation.usage is None
2. No generations produced — When think_max_tokens=0 and answer_suffix=None

These are edge cases that might benefit from unit tests in test/stdlib/sampling/test_think_budget_forcing.py, but they're currently not critical since the budget forcing tests are integration/e2e level (marked with @pytest.mark.e2e, @pytest.mark.qualitative).

Recommendation

The PR can merge without test changes, but to be thorough, you could add a unit test like:
@pytest.mark.unit
def test_think_budget_forcing_no_usage_raises():
"""Verify RuntimeError when backend returns None usage."""
# Would need to mock a backend that returns None for generation.usage

This would ensure the new error paths are exercised, but it's not blocking — existing tests will validate the happy path works correctly.

[ajbozarth]

These are edge cases that might benefit from unit tests in test/stdlib/sampling/test_think_budget_forcing.py, but they're currently not critical since the budget forcing tests are integration/e2e level (marked with @pytest.mark.e2e, @pytest.mark.qualitative).

This did come up when writing but I decided not to since there are no unit tests currently, so I would be introducing them. If including unit tests feels like a blocker I can add them.