feat: complete phase 5 -- error hardening, frontend, Docker, demo, docs

Backend:
- ConversationTracker: Protocol + PostgresConversationTracker for lifecycle tracking
- Error handler: ErrorCategory enum, classify_error(), with_retry() exponential backoff
- Wire PostgresAnalyticsRecorder + ConversationTracker into ws_handler
- Rate limiting (10 msg/10s per thread), edge case hardening
- Health endpoint GET /api/health, version 0.5.0
- Demo seed data script + sample OpenAPI spec

Frontend (all new):
- React Router with NavBar (Chat / Replay / Dashboard / Review)
- ReplayListPage + ReplayPage with ReplayTimeline component
- DashboardPage with MetricCard, range selector, zero-state
- ReviewPage for OpenAPI classification review
- ErrorBanner for WebSocket disconnect handling
- API client (api.ts) with typed fetch wrappers

Infrastructure:
- Frontend Dockerfile (multi-stage node -> nginx)
- nginx.conf with SPA routing + API/WS proxy
- docker-compose.yml with frontend service + healthchecks
- .env.example files (root + backend)

Documentation:
- README.md with quick start and architecture
- Agent configuration guide
- OpenAPI import guide
- Deployment guide
- Demo script

48 new tests, 449 total passing, 92.87% coverage

docs/agent-config-guide.md

@@ -0,0 +1,104 @@
+# Agent Configuration Guide
+
+## Overview
+
+Smart Support agents are defined in `backend/agents.yaml`. Each agent is a
+specialist with a specific role, permission level, and set of tools it can call.
+
+## agents.yaml Structure
+
+```yaml
+agents:
+  - name: order_agent
+    description: "Handles order status, tracking, and cancellations."
+    permission: write
+    tools:
+      - get_order_status
+      - cancel_order
+    personality:
+      tone: friendly
+      greeting: "I can help with your order. What is your order number?"
+      escalation_message: "I'm escalating this to a human agent now."
+
+  - name: refund_agent
+    description: "Processes refund requests."
+    permission: write
+    tools:
+      - process_refund
+      - check_refund_eligibility
+    personality:
+      tone: empathetic
+      greeting: "I'm the refund specialist. How can I help?"
+      escalation_message: "I need to escalate this refund request."
+
+  - name: general_agent
+    description: "Answers general questions and FAQs."
+    permission: read
+    tools:
+      - search_faq
+      - fallback_respond
+```
+
+## Fields
+
+### `name` (required)
+Unique identifier used for routing. Must be alphanumeric with underscores.
+
+### `description` (required)
+Plain-text description of what this agent handles. Used by the supervisor to route
+user messages to the right agent. Be specific.
+
+### `permission` (required)
+Controls the interrupt threshold:
+- `read` -- no interrupt required. Agent can act immediately.
+- `write` -- requires human approval via interrupt before executing tools.
+- `admin` -- requires human approval and is logged for audit.
+
+### `tools` (required)
+List of tool names this agent can use. Tools are registered in the agent factory.
+Each tool name must match a registered LangChain tool.
+
+### `personality` (optional)
+Customizes agent behavior:
+- `tone` -- `friendly`, `formal`, `empathetic`, `technical`
+- `greeting` -- Opening message injected at session start.
+- `escalation_message` -- Message sent when the agent escalates.
+
+## Built-in Templates
+
+Use `TEMPLATE_NAME` environment variable to load a pre-built agent configuration:
+
+| Template | Description |
+|----------|-------------|
+| `ecommerce` | Orders, refunds, shipping, product questions |
+| `saas` | Account management, billing, technical support |
+| `generic` | General-purpose FAQ and escalation |
+
+Example:
+```bash
+TEMPLATE_NAME=ecommerce uvicorn app.main:app
+```
+
+## Adding New Agents
+
+1. Add agent definition to `agents.yaml`.
+2. Register any new tools in `backend/app/agents/`.
+3. Restart the backend.
+
+The supervisor will automatically route to the new agent when the user's intent
+matches the agent's description.
+
+## Agent Routing Logic
+
+1. User sends a message.
+2. The LLM supervisor classifies the intent against all agent descriptions.
+3. If unambiguous, the matching agent is invoked directly.
+4. If ambiguous (multiple plausible agents), the system asks a clarification question.
+5. If multi-intent, agents are invoked sequentially.
+
+## Escalation
+
+Any agent can trigger escalation by calling the `escalate` tool. This:
+1. Sends a webhook notification (if `WEBHOOK_URL` is configured).
+2. Marks the conversation with `resolution_type = escalated`.
+3. Sends the agent's `escalation_message` to the user.