fix(dspy): align span shape with spec, drop over-serialization, avoid nested llm spans

[starfolkai]

Summary

Audit of py/src/braintrust/integrations/dspy/ against .agents/skills/sdk-integrations/SKILL.md.

• Type dspy.module.*, dspy.adapter.format, dspy.adapter.parse, dspy.evaluate as task (were untyped).
• Add metadata.provider on the LM span (derived from LiteLLM-style "openai/..." prefix on instance.model), keep metadata.model.
• Drop over-serialization: removed outputs.toDict() / outputs.__dict__ conversion in on_module_end. Braintrust serializes at send time; DSPy Prediction objects are passed through as-is.
• Fix on_evaluate_end: aggregate eval stats (accuracy, score, total, correct) now go into metadata via an allowlist, not into metrics — the spec restricts metrics to a listed set (tokens, prompt_tokens, time_to_first_token, …).
• Remove redundant current_span().export() parent passing at 5 call sites; start_span picks up the parent from contextvars automatically (matches anthropic/adk).
• Name the LM param whitelist _LM_METADATA_PARAM_ALLOWLIST at module scope (already allowlist-shaped; no denylists found anywhere).
• Do NOT type dspy.lm as llm. DSPy delegates transport to LiteLLM (or a patched provider client), which owns the llm leaf and its token accounting. Typing dspy.lm as llm would produce two nested llm spans per API call when the underlying client is also instrumented — and since DSPy's callback contract doesn't expose usage in outputs, it would also produce a token-less llm span that violates the spec's "tokens required for LLM spans" rule.
• Add a paragraph under Token metrics in sdk-integrations/SKILL.md documenting this pattern for future framework integrations that delegate transport to another instrumentable client.

Test plan

• cd py && uv run nox -s "test_dspy(latest)" — 9/9 passed (dspy 3.2.1)
• cd py && uv run nox -s "test_dspy(2.6.0)" — 9/9 passed
• Extended test_dspy_callback assertions cover the new span_attributes.type values (task on module/adapter spans, not-llm on dspy.lm) and metadata.provider == "openai" on the LM span.
• Existing VCR cassettes reused unchanged.

🤖 Generated with Claude Code

Created by abhijeet

Slack thread