Fix #320: capture prompt + completion content on provider patches

[anilmurty]

The full-request capture (#209) was incomplete for the provider monkey-patches: with capture.prompts = true, a patch_anthropic / patch_openai / patch_gemini / patch_bedrock LLM span got request_params populated but not gen_ai.prompt.content (the messages) or gen_ai.completion.content. Only litellm.py and sdk/agent.py set them. A span without the prompt isn't self-contained enough to replay — the whole point of #209.

Summary

• Added shared record_prompt_content(span, messages) / record_completion_content(span, text) helpers to _request_capture.py, alongside record_request_params/record_request_tools, using the GenAIAttributes.PROMPT_CONTENT / COMPLETION_CONTENT semconv constants (Rule 10).
• record_prompt_content serializes as json.dumps(messages) — the exact shape litellm/sdk/agent.py already use, so a replay harness / backfill reading gen_ai.prompt.content gets one consistent serialization regardless of which patch produced the span.
• Wired both into all four provider patches, set unconditionally on the LLM-call span. strip_captured_content() at ingest already gates them per the [capture] prompts / completions toggles — same pattern litellm uses; no ingest-gate change, no default change (capture stays off by default).

Per provider

• anthropic: messages kwarg + response content text blocks (joined; tool_use blocks skipped). Streaming path captures the prompt; completion there would need stream buffering (out of scope, noted below).
• openai: messages kwarg + choices[0].message.content.
• gemini: contents (first positional arg or kwarg) + response.text, on both the sync and async generate_content patches.
• bedrock: messages/prompt/inputText from the JSON body; completion from the already-parsed response body inside _extract_bedrock_usage (avoids a second read of the response stream).

Also in this PR (the optional refactor)

Refactored litellm.py::_record_prompt_content + the non-streaming completion setter to call the shared helpers, so there's one prompt serialization across every capture path. Dropped the now-duplicate _serialize_prompt (its messages-positional / text-prompt fallbacks are preserved in _record_prompt_content). Behavior-preserving — all existing litellm tests pass.

Tests

tests/unit/test_provider_content_capture.py:

• the shared helpers (incl. json.dumps(messages) shape, None no-ops, non-serialisable fallback);
• each provider's completion/contents extraction (SDK-independent);
• anthropic + openai patches end-to-end through a fake upstream — asserting the span carries PROMPT_CONTENT == json.dumps(messages) and COMPLETION_CONTENT;
• gemini + bedrock via the exact capture calls their patches make (their SDKs aren't installed in CI, so the install path early-returns);
• the ingest gate: both stripped when [capture] is off, kept when on, and gate independently (prompts on / completions off).

Full suite: 1173 passed. ruff + mypy clean on all touched files.

What's NOT in this PR

• Completion content on streaming paths (anthropic/openai stream wrappers): the wrappers don't aggregate the streamed assistant text, so capturing it would mean buffering the stream — a separate concern. The prompt is captured on streaming paths (it's known up front); only streaming completion is deferred.
• Framework patches (langchain / langgraph / crewai / autogen / llamaindex) — explicitly out of scope for this issue.

Closes #320

🤖 Generated with Claude Code

[anilmurty]

Verified against #320, with attention on the one contract that matters most: the serialization. record_prompt_content writes exactly json.dumps(messages), the same shape litellm and sdk/agent already use, with a test asserting it — so a span from any capture path now reads back identically, which is what a replay harness / backfill depends on. Shared helpers in _request_capture.py, semconv constants (Rule 10), wired into all four provider patches, set unconditionally with strip_captured_content gating at ingest (no default change). The litellm consolidation to one serialization is a clean bonus, and the scope discipline is good — streaming completion and framework patches explicitly deferred. Comprehensive tests (helpers, per-provider extraction, anthropic/openai end-to-end, the ingest gate's independent toggles), 1173 passed, ruff + mypy clean, all four CI jobs green.

Ready to merge.