fix(adapter): tolerate missing role in TraceToMessages prompt entries

[itsuzef]

Summary

Some tracers (notably AgentOps when re-emitting prior turns inside gen_ai.prompt.N.*) drop the role key on assistant tool-call entries and tool-response entries. convert_to_openai_messages (agentlightning/adapter/messages.py:134) then crashed with KeyError: 'role' at the first such entry, taking down the entire rollout's adapter pass and silently dropping the trajectory's gradient signal during APO.

This change makes the adapter recover when the role is unambiguous, and otherwise log a warning and skip just the offending message:

• tool_calls present, no role → infer assistant
• tool_call_id present, no role → infer tool
• otherwise → logger.warning(...) once and skip that single entry, preserving the rest of the trajectory's signal

The adapter is the contract between any OTel-emitting tracer and the algorithm layer, so the right place to absorb this kind of upstream variance is here — even if AgentOps tightens its serialization, third-party tracers (Phoenix, hand-rolled agl.emit_*, etc.) will keep producing the same shape.

No API change. Zero new pyright errors against messages.py.

Fixes #425
Fixes #311

Verification

Validated end-to-end inside a linux/amd64 container running the official ghcr.io/astral-sh/uv:python3.12-bookworm image with the project's own lockfile, mirroring the ubuntu-latest CI runners:

uv sync --group dev                                                              # 569 packages installed from uv.lock
uv run --no-sync pytest tests/adapter/ -v                                        # 23 passed (was 21; +2 new regression tests)
uv run --no-sync pyright agentlightning/adapter/messages.py \
                         tests/adapter/test_messages_adapter.py                  # 0 errors, 0 warnings, 0 informations
uv run --no-sync pre-commit run --files agentlightning/adapter/messages.py \
                                       tests/adapter/test_messages_adapter.py    # all hooks Passed (isort, black, prettier, eslint, stylelint, ...)

I also confirmed both new regression tests genuinely fail when the patch is removed, with the exact KeyError: 'role' at messages.py:134 reported in #425 — so the test coverage is meaningful and pinned to the user-reported repro shape.

Test plan

• tests/adapter/test_messages_adapter.py::test_trace_messages_adapter_recovers_assistant_role_from_tool_calls — covers the AgentOps-style re-emitted assistant tool-call entry (no role, has tool_calls) and the matching tool response (no role, has tool_call_id)
• tests/adapter/test_messages_adapter.py::test_trace_messages_adapter_skips_unidentifiable_prompt_entry — covers the unrecoverable case: garbage entry is skipped with a warning, the rest of the trajectory passes through
• Pre-existing tests in tests/adapter/test_messages_adapter.py still pass (no behavior change for well-formed traces)

[itsuzef]

@microsoft-github-policy-service agree