Fix external MCP bridge timeouts with keepalive heartbeats

[justrach]

Summary

• External MCP bridges (Claude Desktop, etc.) impose their own timeout (~60s) on tool calls. Long-running agent tools (run_explorer, run_task, run_reviewer, run_swarm) produce no stdout during execution, so the bridge kills the connection with MCP error -32001 even though the agent is still working fine.
• Adds periodic notifications/message heartbeats every 15s during agent execution. No timeouts imposed — agents run as long as they need.
• Creates src/notify.zig (thread-safe MCP notification sender), modifies runChainStep to run dispatch on a worker thread with heartbeat loop alongside it.

Closes #396, closes #397

Changes

• src/notify.zig (new): Thread-safe notifications/message sender with mutex-protected stdout writes. Supports both line-delimited and Content-Length framing.
• src/tools.zig: runChainStep now spawns agent dispatch on a worker thread and loops on timedWait(15s) sending heartbeat notifications until the agent completes.
• src/main.zig: Wire notify.init(g_use_headers) in the message loop so the notifier picks up the framing mode.

Test plan

• zig build — clean
• zig build test — all pass (including new notify: init sets ready flag test)
• Manual: invoke run_explorer via external MCP bridge, verify heartbeat messages appear every 15s and no -32001 timeout
• Manual: invoke run_task with a long-running preset, confirm it completes without bridge timeout