instrumentation-genai-langchain: ChatOllama produces no inference span (model key is ls_model_name) and tool-call output parts are dropped when the provider reports done_reason

[bytesrabbit]

Version: opentelemetry-instrumentation-genai-langchain 1.1b1
(callback_handler.py:214-227 and :322-388), with langchain-ollama 1.1.0.

What happens.

1. on_chat_model_start resolves the request model from invocation_params and metadata under
   the keys model_name, model_id, model, and returns without a span when none is present.
   ChatOllama publishes the model only as the LangSmith key metadata["ls_model_name"], so no
   inference span is ever produced for it (tool and chain spans are unaffected).
2. on_llm_end builds ToolCallRequest output parts only when the resolved finish_reason is
   tool_calls or tool_use. ChatOllama reports done_reason in generation_info (no
   finish_reason), so the resolved value is "unknown" and the message's tool_calls are
   dropped from gen_ai.output.messages; downstream, tool spans cannot be linked to the
   inference that requested them.

Reproduce. ChatOllama(model=...).bind_tools([tool]).invoke([...]) under the instrumentation.

Suggested fix. Fall back to metadata["ls_model_name"] (and ls_provider, which
normalize_provider already reads) when no model key is present; build tool-call parts whenever
message.tool_calls is non-empty, independent of the provider's finish-reason spelling.

Downstream workaround (forgehawk SDK). ForgehawkLangChainCallbackHandler copies ls_model_name into
metadata["model"] and passes the official on_llm_end a view with finish_reason="tool_calls"
when the message carries tool calls and the official resolution found nothing
(sdk/python/src/forgehawk/instrumentors/langchain.py).

---

Found while building the forgehawk Python SDK (ZioSec) on top of opentelemetry-util-genai; happy to open a PR for the suggested fix if maintainers agree with the direction.