feat: add agent loop integration tests with mocked Claude API

[nickmeinhold]

Problem or Motivation

Current tests cover utilities (sprint, timezone, mentions) and DB queries, but nothing tests the core agent loop pipeline: message in → Claude API call → tool execution → response out. This is where breakage is most likely and most painful.

Proposed Solution

Add integration tests for the agent loop that mock the Anthropic SDK and MCP tool layer:

1. Mock @anthropic-ai/sdk — return canned Claude responses (text-only, with tool calls, multi-round)
2. Mock tool-registry — return canned tool results without real MCP servers
3. Mock grammy Api — capture typing indicators and verify they're sent

Test Scenarios

• Simple text response (no tools)
• Single tool call round (Claude calls a tool, gets result, responds)
• Multi-round tool calls (2-3 rounds of tool usage)
• Tool execution error handling (tool throws, agent recovers)
• Auth failure handling (401 from Claude API)
• Max tool rounds exceeded (10 rounds hit)
• System-initiated messages (no tools available)
• Image attachment handling

Acceptance Criteria

• Integration test file at src/agent/agent-loop.integration.test.ts
• All 8 scenarios above covered
• Tests run in CI without any API keys or external services
• Tests complete in <5 seconds