v0.2.0: WebSocket Streaming

What's new

• Multi-stage WebSocket streaming via /api/ws
• Query cancellation mid-stream with partial answer preserved in audit
• Real-time status events: retrieving → generating → verifying_trust
• TTFT: 3-8s → <500ms (measured in E2E tests)
• Groq rate limit (429) → clear GROQ_RATE_LIMIT error with retry_after_ms
• Exponential backoff auto-reconnect (1s → 2s → 4s → 8s cap)
• uvicorn WS ping/pong keepalive (20s interval)

Test coverage

• 14 backend tests (unit + E2E)
• 12 frontend tests (hook state machine + stale ID filtering)
• ruff lint clean

Why this matters

Users previously waited 3-8s staring at a spinner. Now they see first token in <500ms and can cancel expensive queries mid-stream.

Phase 1 tasks

• P1-1: Backend WS endpoint + connection manager
• P1-2: QueryTask async state machine with cancellation
• P1-3: Groq streaming integration + error handling
• P1-4: Frontend useWebSocketQuery hook + reconnect
• P1-5: Streaming UI components with multi-stage status
• P1-GATE: E2E verification + metrics + this release