Review request: fix(acp) steer channel busy/stale-ack + Git-bash Windows test harness (PR #4043)

## Context

This issue references PR #4043, which fixes three bugs in the `buzz-acp` crate:

1. **Don't cancel a healthy turn when the steer channel is busy** — a burst of mentions filled the capacity-1 steer channel; `TrySendError::Full` was treated as a transport error, firing the universal cancel+merge fallback and destroying the in-flight turn. `send_steer` now maps `Full` to `SteerError::SteerChannelBusy`, `try_native_steer` returns a tri-state, and on Busy the event stays queued for normal dispatch.
2. **Skip cancel+merge fallback for a stale steer ack** — a late `SteerAck` targeting an already-completed turn could cancel a healthy newer turn. The ack now carries the steer-time `turn_id`; stale acks are skipped with a `tracing::warn!`.
3. **Spawn harness scripts on Git bash, not the WSL shim** — `Command::new("bash")` on Windows resolved the WSL launcher instead of Git-for-Windows bash, causing deterministic steer-capture test failures and the 18-24 test flake bucket. The harness now resolves a Git bash explicitly.
## Verification

- `cargo test -p buzz-acp --lib` — 663/663 pass on the rebased branch
- - `cargo clippy --workspace` — clean
- - - `cargo build` (workspace) — clean
- - - - Branch rebased onto current `main` (`b1b283cd4`)
## Request

All commits are DCO-signed; the three reachable checks (Semgrep OSS, zizmor, DCO) pass. The PR is blocked on a maintainer approving the fork-PR workflow runs and providing a code-owner review. Could a maintainer please take a look? Thanks!