Step 1 of M3.5: the shared mechanism. M3.4 and M3.5 build on it.
- Engine: `Outcome.pending: Option<Pending>`. `Pending::new(rx, timeout,
finish)`; `Pending::wait()` is called without the engine lock and
returns at the answer, a dropped sender or the deadline (timeout + 1 s).
`Engine::resolve(answered)` then runs `finish(state, reply)` inside the
guard, under the lock. `run_waiting(lock, cmd, params, publish)` does
all three steps for hosts.
- Hosts: `Host::run`, `testkit::EngineBackend` and the control test
backend use `run_waiting`.
- New IPC `engine_execute_waiting` (`#[tauri::command(async)]`, off the
main thread) with `runWaiting()` in `ui/src/ipc.ts`. Only `lsp.*`
requests use it.
Ordering guarantee:
- `engine_execute` stays synchronous on the main thread, so ordinary
commands (`buffer.edit`) still run in call order.
- The order towards the server is fixed under the engine lock:
`buffer.edit` queues its didChange on the session's FIFO queue before
it returns, and `lsp::submit` queues the request on the same queue.
A request therefore sees every edit whose `buffer.edit` returned
first, and the UI awaits the sync's `idle()` before sending one.
Failures:
- A stall is caught by the request timeout, with the deadline as
backstop.
- On restart, close, quit or crash the request is dropped or resolved
as stopped, so the waiter gets an error, never a hang.
Tests: `crates/engine/tests/pending.rs` (answer, lock free while
waiting, dropped sender, deadline, resolve, lsp.restart while a
request is pending) and Vitest for `runWaiting`. Gate: 13 steps,
321 s, Vitest 443 + 1 skipped.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>