feat(llm): MLflow tracing for the Agent SDK subprocess (migration, phase 2)

[PalmPalm7]

What

The Agent SDK (#24) runs the Claude Code CLI as a child process, so the legacy app-side MLflow collector (src/connections/mlflow_tracking.py) never sees its LLM calls. This adds src/llm/sdk_tracing.py to trace the SDK path the way RHDP already traces Claude Code itself — by environment — so the subprocess exports its own claude_code.* spans (tokens, per-turn latency, tool calls) to the same MLflow server/experiment.

Why

Phase-2 needs trustworthy per-call cost/latency/tool numbers to benchmark the SDK path against legacy. This is the measurement layer.

How

• build_tracing_env(config) derives MLFLOW_CLAUDE_TRACING_ENABLED / MLFLOW_TRACKING_URI / MLFLOW_EXPERIMENT_NAME (+ basic-auth) from the existing mlflow.* config — the same reads as init_mlflow. Returns {} when tracking_url is empty, so it's a no-op by default, exactly like the collector.
• AgentSdkClient.from_config merges that env into the SDK subprocess env automatically; an explicit agent.sdk.env wins on conflict.
• build_hooks_settings() returns the settings.json hooks block (Stop + SessionStart) for deployments that prefer hook-based wiring.

Scope / safety

Additive; only affects the sdk runtime, and only when mlflow.tracking_url is set. No change to the legacy path or default deploy.

How to test

pytest tests/test_sdk_tracing.py -q   # 9 tests

Result (local gate)

• black ✓ · ruff ✓ · mypy ✓
• pytest tests/test_sdk_tracing.py → 9 passed; full suite → 96 passed, no regressions
• End-to-end span export against the live MLflow parsec-agent-metrics is pending in-cluster verification — results will be commented below once run.

Builds on #24 (SDK adapter) and #30 (AgentRunner). Plan: artifacts/parsec-agent-sdk-migration-plan.md.

[PalmPalm7]

Test results.

• Local gate: black ✓ · ruff ✓ · mypy ✓ · pytest tests/test_sdk_tracing.py → 9 passed; full suite 96 passed, no regressions.
• Upstream CI: quality-gates + docker-build + ci-status all green.
• End-to-end span export to the live MLflow (parsec-agent-metrics) is pending in-cluster verification — the personal NERC cluster is currently at pod/disk capacity (no schedulable nodes). Will post the trace evidence here once capacity frees.

[rut31337]

Code Review — PR #31 (Draft)

Scope: 3 files, +214 lines — MLflow tracing for Agent SDK subprocess
Effort: High

Findings (2)

1. build_hooks_settings() is dead code (src/llm/sdk_tracing.py:80)
Defined but never called anywhere in the codebase. Adds maintenance surface with no current value. Consider removing until there's a concrete caller.

2. Duplicated _section() helper (src/llm/sdk_tracing.py:97)
Identical to _get_section() in agent_sdk_client.py. Bug fixes must be applied to both. Extract to a shared location — both modules already live under src/llm/.

No blocking issues. Core tracing logic is correct — build_tracing_env() properly derives MLflow env vars from config, no-op path returns {} when tracking_url is empty, and env merge precedence in from_config is correct. Test coverage is solid across 9 tests.