fix(llmobs): skip downstream OpenAI span check for litellm_proxy/ models

[ZStriker19]

Problem

The ddtrace litellm integration produces empty LLMObs spans when using a litellm_proxy/ model (e.g. litellm_proxy/azure-gpt-5-nano). The request routes through a proxy, so the OpenAI integration never fires — but the span was being suppressed as if it would.

Root Cause

_has_downstream_openai_span() does a substring check for "azure", "gpt", "openai" in the model name. A model like litellm_proxy/azure-gpt-5-nano matches "azure", so submit_to_llmobs=False is set and the span is dropped entirely.

Fix

Short-circuit in _has_downstream_openai_span() when the model name starts with litellm_proxy/ — these requests always go through a proxy, never directly to OpenAI/Azure.

Testing

Added 9 parametrized unit tests for _has_downstream_openai_span() covering: proxy suppression (the fix), normal OpenAI/Azure passthrough, streaming, disabled OpenAI integration, and non-OpenAI models. No VCR cassette needed — the method is pure logic.

Closes MLOB-6958

🤖 Generated with Claude Code

[cit-pr-commenter-54b7da]

Codeowners resolved as

ddtrace/llmobs/_integrations/litellm.py                                 @DataDog/ml-observability
releasenotes/notes/litellm-proxy-model-llmobs-fix-a1b2c3d4e5f6a7b8.yaml  @DataDog/apm-python
tests/contrib/litellm/test_litellm_llmobs.py                            @DataDog/ml-observability