fix(runtime): handle trigger() rejections when waking idle agents

[guo-yu]

Problem

AgentProcess.trigger() is an async method that can reject (it awaits relay.recv, credentialManager.ensureFresh(), runOneShot(), and the preamble/env providers). The supervisor called it fire-and-forget — without await or .catch() — in two places:

1. Inside the setNewMessageCallback lambda that wakes idle on-demand / idle_cached agents when a message arrives.
2. Inside wakeUnblockedAgents() when a completed task unblocks dependent agents. The surrounding try/catch there does not cover the promise, since it is never awaited.

Any rejection became an unhandled promise rejection, which can crash the whole supervisor process.

Fix

Both call sites now use:

void agent.trigger().catch((err) => {
  log.error('agent trigger failed', {
    agent: agentName,
    error: err instanceof Error ? err.message : String(err),
  });
});

This matches the existing agent.start().catch(...) error-handling pattern already used in Supervisor.start().

Tests

• Added two regression tests asserting that no unhandledRejection event is emitted when a woken agent's trigger() rejects, covering both the new-message wake path and the task-unblock wake path. Both tests fail without the fix and pass with it.
• The tests use plain function mocks rather than vi.fn(), because vitest mocks attach internal handlers to returned promises (for mock.settledResults), which would mask the unhandled rejection.
• Also tightened an existing loose mock (trigger: vi.fn() returned undefined, which is not a valid Promise<void>).

Full @wanman/runtime suite: 48 files, 716 passed / 8 skipped (pre-existing skips). tsc --noEmit clean.

🤖 Generated with Claude Code