refactor(telemetry): migrate application spans to plugin hooks

[ajbozarth]

Pull Request

Issue

Fixes #1048. Phase 2 of #444.

Description

Action span emission moves from inline trace_application calls in stdlib/functional.py to a new ComponentTracingPlugin subscribed to component_* hooks. Session and start_session spans stay in stdlib but route through clean typed helpers in tracing.py instead of the deprecated trace_application surface. Hooks fit spans whose lifecycle lives inside one async function (one Task); session lifecycle crosses multiple _run_async_in_thread calls and OTel Token attach/detach is task-affine, so direct emission is the right shape there.

Reviewer call-outs:

1. _run_async_in_thread now propagates the calling thread's contextvars into the new Task. One-way, no copy-back. Without it, asyncio.run_coroutine_threadsafe starts the new Task with empty contextvars, so contextvar-backed state on the user thread (OTel active span, Mellea's session_id / request_id / model_id / sampling_iteration, log context) is invisible inside Mellea's async work — including the session > action > chat hierarchy that the docs have promised since feat: instrument telemetry #355. With this PR, that hierarchy actually forms in real usage for the first time. ~6 lines in event_loop_helper.py; aligns Mellea with how asyncio.create_task already inherits Context within a thread.

2. Span renames. session_context → session; aact → action. session_context's suffix was redundant, and aact named a Mellea internal function rather than the operation. Trace-tooling consumers filtering on the old names need to update.

3. Scope-narrowing change vs. the original issue. Issue feat: session and act spans via plugin hooks #1048 asked for session and act spans via plugin hooks. Action lands as planned (ComponentTracingPlugin); session does not. Session lifecycle crosses multiple _run_async_in_thread calls (sync __enter__, sync m.act(...), sync __exit__/cleanup()), each scheduled as a separate Task with its own contextvars.Context. OTel Token.reset is bound to the originating Context, so a span attached in one hook's Task can't be detached in another. Session spans now emit directly from MelleaSession.__enter__/__exit__ and start_session() via typed helpers in tracing.py. The SESSION_* hooks still fire as observation points for non-tracing plugins.

4. Component payload component_id. Adds a UUID correlation field on ComponentPreExecutePayload / ComponentPostSuccessPayload / ComponentPostErrorPayload, matching the existing generation_id pattern on generation payloads. Plugin authors subscribing to component_* hooks gain a correlation key.

5. Bug fix from refactor(telemetry)!: move backend tracing onto plugin/hook pattern #1181 review. mellea.response is now gated on MELLEA_TRACES_CONTENT / OTEL_INSTRUMENTATION_GENAI_CAPTURE_MESSAGE_CONTENT. Pre-this-PR it was captured unconditionally; this is the first real caller of is_content_tracing_enabled() (per @planetf1's refactor(telemetry)!: move backend tracing onto plugin/hook pattern #1181 review).

6. Streaming out of scope. mellea/stdlib/streaming.py was added in PR feat(stdlib): add streaming event types, events() iterator, and OTEL bridge (#902) #1095 mid-epic, after the original Phase 2 plan was set, and it imports the deprecated trace_application surface. Migrating it is tracked as a follow-up Phase 2 sub-issue feat(telemetry): migrate stream_with_chunking orchestration span to plugin hooks #1290. The four deprecated public helpers (trace_application, set_span_attribute, set_span_error, set_span_status_error) stay in tracing.py until streaming migrates off them.

7. Docs touched only where existing claims went stale. Span-hierarchy diagrams updated for the renames; one stale Phase-1 entry corrected (mellea.backend is a string identifier, not a class name). Broader docs cleanup is tracked in docs: add Telemtry example for getting full traces #945.

8. Bug fix: mellea.sampling_success now reflects the actual flag. Pre-this-PR the attribute was bool(sampling_result.result) — i.e. truthiness of the chosen ModelOutputThunk, which is True whenever sampling produced any non-None output even if no requirement passed. Post-this-PR it reads sampling_result.success, the flag the field name has always promised. Traces where sampling exhausted retries without satisfying requirements will now correctly show mellea.sampling_success=false.

BREAKING CHANGE: Span names renamed:

• session_context → session
• aact → action
  Trace-tooling consumers filtering on the old names need to update.

Testing

• Tests added to the respective file if code was changed
• New code has 100% coverage if code was added
• Ensure existing tests and github automation passes (a maintainer will kick off the github automation when the rest of the PR is populated)

Attribution

• AI coding assistants used

---

Adding a new component, requirement, sampling strategy, or tool?

If your PR adds or modifies one of the types below, check the matching box. A checklist of type-specific review items will be posted as a comment.

• Component
• Requirement
• Sampling Strategy
• Tool

NOTE: Please ensure you have an issue that has been acknowledged by a core contributor and routed you to open a pull request against this repository. Otherwise, please open an issue before continuing with this pull request.