Managed agents should follow the user's threading choice

## Problem

Managed-agent replies currently open a thread beneath every top-level human message. This overrides the user's conversation layout and makes a channel harder to scan when the user intentionally writes at the top level.

## Expected behavior

- A top-level human message receives a top-level agent reply.
- A human message within a thread receives a reply anchored to that thread's root, preserving the existing flat layer-1 behavior.
- Agent-to-agent turns retain their existing freedom to nest when useful.

The user's placement of the triggering message should determine the agent reply placement.

## Proposed implementation

Update the reply-anchor resolution and prompt contract in:

- `crates/buzz-acp/src/queue.rs`
- `crates/buzz-acp/src/base_prompt.md`

For human-facing turns, return a reply anchor only when the triggering event already belongs to a thread. Keep the existing thread-root flattening and agent-to-agent behavior unchanged.

## Verification

Add or update regression coverage for top-level human messages, threaded human messages, mixed batches, DMs, and agent-to-agent turns, then run:

```sh
cargo test -p buzz-acp --lib
cargo clippy -p buzz-acp --all-targets -- -D warnings
cargo fmt --check
git diff --check
```

Start in crates/buzz-acp/src/queue.rs and crates/buzz-acp/src/base_prompt.md, tracing reply-anchor resolution and the prompt contract. Add regression coverage for top-level and threaded human messages, mixed batches, DMs, and agent-to-agent turns, then run the listed cargo test, clippy, fmt, and git diff checks.

Written by the indexing model from the issue text.