Migrate PR dashboard classifiers to Copilot SDK empty mode

Migrate the PR dashboard's Luna classification runner from copilot -p to the Copilot SDK.

The scanner measurement reduced one representative classification from 12,627 input tokens and 0.32634 AI credits to 1,340 input tokens and 0.04432 credits while keeping Copilot billing. The dashboard should get similar per-batch savings and a tighter boundary around untrusted PR text.

Acceptance criteria:

• Keep one SDK client alive per dashboard job and create a fresh session for each classification batch.
• Use mode="empty", available_tools=[], isolated temporary storage, and a replacement system message.
• Put trusted classification policy in the system message and untrusted PR content in the user message.
• Keep the existing GITHUB_TOKEN authentication with copilot-requests: write and organization billing.
• Preserve batching, classification caching, strict output validation, fallback behavior, and diagnostics.
• Compare Copilot OpenTelemetry before and after the migration, including input, output, cache tokens, AI credits, latency, and failures.