openai: chat message content recorded as empty parts when it is a list of content parts

[HQidea]

Package: opentelemetry-instrumentation-genai-openai (current main; the same code ships in opentelemetry-python-contrib as opentelemetry-instrumentation-openai-v2 ≤ 2.4b0, which per that repo's guidelines now only takes security patches — hence filing here).

Bug

When a Chat Completions message's content is the list-of-content-parts form (the standard OpenAI shape for multi-part text and for multimodal requests), the instrumentation records the message with empty parts in gen_ai.input.messages:

{"role": "user", "parts": []}

The message content is silently dropped from telemetry, even though content capture is enabled and the actual request carried the content.

Reproduction

from opentelemetry.instrumentation.genai.openai.utils import _prepare_input_messages

messages = [
    {"role": "system", "content": "You are a fact extractor."},
    {
        "role": "user",
        "content": [
            {"type": "text", "text": "user: remember, my name is Tom"},
            {"type": "text", "text": "assistant: got it, Tom!"},
        ],
    },
]
for m in _prepare_input_messages(messages):
    print(m.role, len(m.parts))
# system 1
# user 0    <-- content dropped

End-to-end this shows up as GENERATION observations with an empty user message in any OTLP backend that renders gen_ai.input.messages (we hit it in production tracing multi-part-text extraction calls into Langfuse and spent a while chasing a phantom "empty prompt" bug in our own pipeline).

Root cause

_prepare_input_messages gates content on _is_text_part, which only accepts str (or an iterable of str):

def _is_text_part(content: Any) -> bool:
    return isinstance(content, str) or (
        isinstance(content, Iterable)
        and all(isinstance(part, str) for part in content)
    )

A content-parts array is a list of dicts (or typed part objects), so it fails the check and nothing is appended. _prepare_output_messages shares the same helper.

Note the semconv message models already have the right part types for these shapes (Text, Uri, Blob, File in opentelemetry-util-genai), and the sibling opentelemetry-instrumentation-genai-anthropic package already converts content blocks per-part (convert_content_to_parts) — the OpenAI package just predates that treatment.

Expected

• {"type": "text", ...} parts → one Text part each
• {"type": "image_url", ...} → Uri (modality image)
• {"type": "input_audio", ...} → Blob (modality audio)
• {"type": "file", ...} → File
• unrecognized part types skipped rather than nuking the whole message

I have a fix with unit tests ready and will open a PR referencing this issue.