feat: wire LLM cost tracking end-to-end — backend, API, WebSocket, and CostDashboard UI

[nuthalapativarun]

Feature: LLM Cost & Token Usage Tracking

This is PR 2 of 2. Depends on #319 which adds LLMCallEvent and CostThresholdExceededEvent to the event bus. Merge #319 first.

UFO already computes cost per LLM call in ufo/llm/base.py and returns it from every provider — but the value is discarded by callers. This PR wires that existing data all the way from the LLM call layer through the event bus, into a metrics observer, out through a FastAPI REST API and WebSocket, and into a live-updating React dashboard in the Galaxy Web UI.

---

What this PR does

Backend — LLM call layer (ufo/llm/)

• base.py: get_cost_estimator() now returns CostResult(cost, prompt_tokens, completion_tokens) (a NamedTuple) instead of a bare float. Backward-compatible — callers that only used .cost still work.
• openai.py, claude.py, gemini.py: Updated to unpack CostResult from get_cost_estimator().
• llm_call.py: After each provider call, measures wall-clock duration and emits LLMCallEvent on the Galaxy event bus. Falls back gracefully (logs at DEBUG) if Galaxy is not running.

Backend — Observer (galaxy/session/observers/base_observer.py)

• SessionMetricsObserver now subscribes to LLM_CALL_COMPLETED events and accumulates:
  • Running totals: cost, prompt tokens, completion tokens, API call count
  • Per-agent and per-model cost breakdowns
  • Rolling call log (capped at last 500 entries)
• Configurable cost_alert_threshold (default 0.0 = disabled): emits CostThresholdExceededEvent once when the session total exceeds the threshold.

Backend — API (galaxy/webui/)

• models/responses.py: Adds LLMCallRecord and SessionCostSummary Pydantic response models.
• services/metrics_service.py: Thin service that reads from SessionMetricsObserver.metrics["llm_metrics"] and serialises to JSON or CSV.
• routers/metrics.py: New FastAPI router:
  • GET /api/metrics/cost — returns SessionCostSummary for the active session
  • GET /api/metrics/cost/export?format=json|csv — downloads the full call log
• Registered in server.py alongside existing routers.

Backend — WebSocket (galaxy/webui/websocket_observer.py)

• EventSerializer extended to handle LLMCallEvent (broadcast as message_type: "llm_metrics_update") and CostThresholdExceededEvent (broadcast as message_type: "cost_alert").

Frontend — Zustand store (src/store/galaxyStore.ts)

• New LLMMetrics interface and LLMCallRecord type.
• llmMetrics slice with setLLMMetrics and appendLLMCall actions. appendLLMCall incrementally updates all aggregates client-side (no full-fetch needed on each event).
• rightPanelTab union extended with 'cost' tab.

Frontend — WebSocket handlers (src/main.tsx)

• handleLLMMetricsUpdate: maps llm_metrics_update WebSocket messages to appendLLMCall.
• handleCostAlert: maps cost_alert messages to a pushNotification with severity: "warning".
• GalaxyEvent interface extended with LLM-specific fields (removes all as any casts).

Frontend — UI (src/components/metrics/)

• CostByModelChart.tsx: Tailwind-only horizontal bar chart, sorted by cost descending. Reused for both model and agent breakdowns.
• CostDashboard.tsx: Live-updating panel showing:
  • Summary row: total cost, prompt tokens, completion tokens, API call count
  • Cost by model bar chart
  • Cost by agent bar chart
  • Collapsible recent-calls table (last 50, newest first)
  • Dismissible amber alert banner on cost_alert notifications
• RightPanel.tsx: Adds a tab bar ("Constellation" / "Cost") to switch between the existing constellation view and the new cost dashboard.

---

Architecture

LLM Provider (openai / claude / gemini)
  └─ CostResult(cost, prompt_tokens, completion_tokens)
       └─ llm_call.py → emits LLMCallEvent on EventBus
            └─ SessionMetricsObserver
                 ├─ accumulates llm_metrics dict
                 └─ emits CostThresholdExceededEvent (if threshold set)
                      └─ WebSocketObserver → broadcasts to frontend
                           ├─ "llm_metrics_update" → appendLLMCall (Zustand)
                           └─ "cost_alert"         → pushNotification (Zustand)
                                └─ CostDashboard (React) — live updates
GET /api/metrics/cost         → SessionCostSummary (REST poll fallback)
GET /api/metrics/cost/export  → JSON or CSV download

---

Testing

• All existing tests pass (no breaking changes — cost float still returned from get_completions()).
• GET /api/metrics/cost returns 404 when no session is active; 200 with live data during a session.
• WebSocket cost updates verified by running a session and observing the CostDashboard panel update in real time.