fix(buzz-acp): clear steer_rx on all run_prompt_task exit paths

[wpfleger96]

Problem

Any first-turn error in run_prompt_task — session-create failure, initial_message error, etc. — returns the agent via PromptResult before the read loop's steer_rx.take() runs. This leaves steer_rx = Some on the returned agent. The next dispatch for the same channel reclaims the same AcpClient, install_steer_rx hits its is_none() assert, and the main thread panics — far from the original error.

Root cause traced to pool.rs: there are ~22 early-return paths in run_prompt_task, all of which send PromptResult { agent, .. } back to the pool. The only steer_rx.take() is at acp.rs:890 inside the read loop, which none of the error paths reach.

Fix

Add AcpClient::clear_steer_rx() (idempotent, sets steer_rx = None) in acp.rs.

Add a send_prompt_result() helper in pool.rs that calls agent.acp.clear_steer_rx() immediately before moving agent into PromptResult. Every exit path in run_prompt_task is routed through this helper — making the install_steer_rx invariant structurally unviolatable rather than a reachable crash. The assert! in install_steer_rx is preserved as a true never-happens guard.

Tests

Two new unit tests in pool::tests:

• test_send_prompt_result_clears_steer_rx_on_early_return: installs a receiver, calls send_prompt_result without the read loop running (early-return simulation), asserts steer_rx is None and a subsequent install_steer_rx does not panic.
• test_send_prompt_result_is_noop_when_steer_rx_already_consumed: receiver already None (happy path), asserts clear_steer_rx is idempotent and install_steer_rx still does not panic.

Files changed

• crates/buzz-acp/src/acp.rs: clear_steer_rx() method + steer_rx_is_none() test helper
• crates/buzz-acp/src/pool.rs: send_prompt_result() helper + 22 call-site replacements + 2 new tests