refactor(dspy): use acall usage tracking

[enitrat]

Summary

• Switch DSPy module calls to and rely on prediction usage tracking
• Replace manual aggregation with tracker for streaming usage
• Update tests/mocks for and optional optimizer artifacts

Context

DSPy attaches usage in module wrappers; moving to avoids missing usage and centralizes tracking.

Test Plan

• PYTHONPATH=python .venv/bin/python -m pytest -q python/tests

[enitrat]

[AUTOMATED]

Summary

This PR refactors DSPy module invocations from aforward to acall to leverage DSPy's built-in usage tracking mechanism, and replaces manual token usage accumulation with dspy.track_usage() context manager for streaming. This is a solid architectural improvement that simplifies code and improves reliability.

Strengths

• Cleaner usage tracking: Removing manual _accumulate_usage and relying on DSPy's acall / track_usage() is the right approach
• Reduced complexity: ~11 fewer lines of production code with less manual plumbing
• Better separation of concerns: Usage tracking is now handled by the framework where it belongs
• Comprehensive test updates: All mocks properly updated to include acall alongside existing methods

Minor Observations

• The combine_usage function in types.py is no longer imported in rag_pipeline.py but remains in the codebase (used elsewhere or could be dead code)
• The optimizer_artifacts_optional fixture globally sets OPTIMIZER_RUN=1 which is a reasonable approach for test isolation

Overall, this is a well-executed refactor that improves code quality. Tests pass according to the PR description.