feat(mcp): Include classification label in `logs` tool response

[jaroslawgajewski]

Problem

gh-aw performs risk classification on each run (risky / normal / baseline) as
part of the detection pipeline. This classification drives the conclusion step
behavior (approve / request-changes) and is visible in the gh-aw.

However, the classification label is not included in the agentic-workflows
MCP logs tool response. Any consumer of the logs tool — observability
exporters, audit workflows, trend analysis — cannot access this data.

Current behavior

The logs response includes tokens, cost, duration, and trigger type,
but classification is absent. Consumers that try nested key paths
(comparison.classification.label) get null/undefined.

Expected behavior

Each episode should include:

Field	Type	Values
classification	string	"risky", "normal", "baseline", "unclassified"

Use cases unlocked

• Risk monitoring: Track risky-vs-normal run distribution over time.
• Alerting: Trigger alerts when risky run percentage exceeds a threshold.
• Model selection validation: Correlate classification with model tier to
  validate that models produce acceptable risk levels.
• Observability: Any external dashboard can filter and trend by risk level.
• Audit compliance: Evidence that detection is classifying runs correctly.

[pelikhan]

@lpcox @davidslater