Kimi harness launches `kimi --acp` (unknown option) and all Kimi agents fail to start

**Summary**
The built-in **Kimi** runtime invokes the Kimi Code CLI as `kimi --acp`, but the current Kimi Code CLI expects the subcommand `kimi acp` (no leading dashes). Every Kimi-harness agent crashes on launch with "unknown option '--acp'", and buzz-acp reports "all N agents failed to start — cannot continue".

**Environment**
- Buzz hosted relay (communities.buzz.xyz), desktop app, macOS
- Kimi Code CLI at ~/.kimi-code/bin/kimi (MoonshotAI/kimi-code)
- Built-in "Kimi" harness (id `kimi-code`); Arguments field = `--acp`

**Steps to reproduce**
1. Install Kimi Code CLI and log in (`kimi`, then `/login`).
2. Create an agent using the **Kimi** harness.
3. Start the agent.

**Expected**
buzz-acp launches Kimi in ACP mode via `kimi acp`; the agent initializes.

**Actual**
buzz-acp launches `kimi --acp`; Kimi exits and every instance fails:

agent_cmd=/Users/.../.kimi-code/bin/kimi --acp ...
error: unknown option '--acp'
ERROR buzz_acp: agent initialize failed: Agent process exited unexpectedly agent=0
Error: all N agents failed to start — cannot continue

**Workaround**
Point the Kimi harness Command at a shim that ignores the flag and runs the subcommand:

#!/usr/bin/env bash
exec "$HOME/.kimi-code/bin/kimi" acp

**Suggested fix**
The Kimi harness should invoke `kimi acp` (subcommand), not `kimi --acp`. Ref: Kimi Code docs — the `kimi acp` subcommand.

Start at the built-in Kimi harness configuration and its Arguments field, then trace the buzz-acp agent launch entry point that produces the shown agent_cmd. Change the Kimi invocation to use the `acp` subcommand and verify that a Kimi agent initializes without the unknown-option error.

Written by the indexing model from the issue text.