fix(engine): bind shared cancel slot to the owning turn

[asto18089]

Fixes the foundation half of Pinvou/pinvou-agent#254.

Problem

Hosts cancel through a shared token slot that carried no turn identity: EngineHandle::cancel_with_mode fired whatever token occupied the slot at call time. For host-driven turns the host-side generation gate made that safe — the host reserves the next turn (advancing its epoch under the same lock the cancel checks) before the engine can start one.

The three runtime self-start paths — idle sub-agent completion, background shell completion wake, and goal continuation — call handle_send_message inside the engine and swap the shared token before any host reserve. A host cancel whose generation view was still on the finished turn passed the epoch check and fired the follow-up turn's token: the C1 issued to cancel turn N killed the engine's self-started turn N+1.

Fix

• The shared slot becomes TurnCancelSlot { turn_id, token }, swapped atomically at every turn start. handle_send_message and the user ! shell-command turn mint the turn id first, then install the token under that identity.
• New host entry point EngineHandle::cancel_turn(turn_id, reason, mode) -> bool. The identity check and the token clone happen under the same slot lock the install uses, so the decision is atomic against a concurrent turn start, and the cloned token can only ever fire the named turn. On identity mismatch nothing is cancelled and no steer disposition or cancel reason is published — the target turn is already gone.
• cancel_with_mode keeps its exact fire-current-token semantics for single-user frontends (TUI stop button, Op::CancelRequest).

Tests

• engine_handle_cancel_turn_only_fires_the_named_turns_token — slot contract: unnamed slot skips, stale id skips without firing the follow-up turn's token, matching id fires.
• forkguard_cancel_turn_binding_spares_unnamed_turns_and_hits_the_observed_turn — regression through a real engine with an injected blocking model client: a foreign turn id leaves the in-flight request untouched (is_cancelled false, provider future not dropped), the id observed from TurnStarted interrupts exactly that turn with the usual TurnComplete(Interrupted).

cargo test -p codewhale-tui --lib core::engine:: (504) and forkguard (58) slices pass; full suite runs in the paired parent-repo PR.

The paired parent-repo PR wires EnginePool::cancel and the forwarder's pending-cancel replay to this API.

[github-actions]

Thanks @asto18089 for taking the time to contribute.

This repository is observing a maintainer-managed PR intake gate in dry-run mode, so this pull request is staying open. This note helps maintainers prepare the allowlist before any enforcement is considered.

Please read CONTRIBUTING.md for the expected contribution shape. A maintainer can grant recurring PR access by commenting /lgtm on a pull request.