SP-5/SP-8: EMA cost model, dynamic scraper, backend delivery

[zzylol]

Summary

• SP-5 EMA cost model: online_cost_model wired into CostModelPlanner via with_online_store(). Scoring now uses a live-blended cost table (70 % online / 30 % benchmark after 5+ observations) instead of static benchmarks, so the planner converges toward real-world behaviour over time.
• SP-8 feedback loop: Scraper gains dynamic endpoint management (add_endpoint, remove_endpoint, set_sketch_type) backed by Arc<RwLock<Vec<Endpoint>>>. New on_metrics callback delivers ScrapedData (sketch_size_bytes + derived cpu_micros_per_sample) to the EMA store after each successful scrape.
• OpAMP role tracking: X-Agent-Role: agent|backend header distinguishes edge collectors from aggregation backends. push_to_role() delivers role-appropriate configs: agent YAML to Agent role, backend YAML to Backend role. on_connect/on_disconnect hooks auto-register/deregister scrape endpoints.
• main.rs wiring: initialise EMA store → build Scraper with EMA callback → build OpampServer with connect/disconnect hooks → CostModelPlanner::with_online_store → start scraper background loop. After each plan: push agent config to agent-role collectors, backend config to backend-role collectors, update endpoint sketch types for EMA attribution.

Test plan

• cargo test — all 128 tests pass
• monitor — 8 tests including new on_metrics_callback_fires and add_remove_endpoint
• opamp — role parsing, push_to_role, connect/disconnect
• planner::online_cost_model — EMA blending, store init, effective_table fallback
• All existing cost model, delta cost model, query parser, analyzer, config, and store tests continue to pass

🤖 Generated with Claude Code