Improve EventBus reliability and error observability

[longzhi]

Problem

Two related reliability problems in the event bus system:

1. Silent message dropping

clawhive-bus/src/lib.rs:86 — try_send() with buffer size 8. When subscribers are slow, messages are silently dropped. Critical messages like NeedHumanApproval could be lost without any indication.

2. Ignored publish failures

clawhive-core/src/shell_tool.rs:81,133 — let _ = bus.publish(...) ignores publish failures with no logging. If a publish fails, there's no way to know.

Proposed Solution

1. Increase channel buffer size or implement backpressure strategy for critical message types
2. Add tracing::warn! on publish failures across the codebase
3. Consider differentiating critical vs. best-effort messages

Impact

• Reliability — prevents silent loss of critical messages (e.g., human approval requests)
• Observability — publish failures become visible in logs for debugging