"Winnable deals only" used to call `choose_winnable_seed` inline on
the main thread inside `handle_new_game`. Each rejected attempt costs
~120 ms (`SolverConfig::default()` budget); the loop caps at
`SOLVER_DEAL_RETRY_CAP` = 50, so a pathological run could stall the
UI for ~6 s on a New Game click. Quat flagged this as the highest-
impact UX regression left in the engine.
Reorganised so the solver runs on `AsyncComputeTaskPool`:
- New `PendingNewGameSeed` resource holds an `Option<PendingSeedTask>`
carrying the in-flight `Task<u64>` plus the request's `mode` and
`confirmed` flags so the polling system can replay them on a
synthetic `NewGameRequestEvent` once the task resolves.
- `handle_new_game` now writes to that resource (and `continue`s)
for the winnable-only / Classic / random-seed branch, instead of
calling `choose_winnable_seed` synchronously.
- `poll_pending_new_game_seed` runs `.before(GameMutation)` so the
synthetic event lands in the same frame's `handle_new_game` —
the player sees no extra-frame visual lag once the solver
completes.
- Cancel-on-replace: when a fresh `NewGameRequestEvent` arrives
while a previous task is in flight, `pending_seed.inner = None`
drops the old task (Bevy's `Task` Drop cancels cooperatively at
the next await point) before processing the new request.
Two tests:
- `winnable_seed_search_runs_async_and_completes_eventually` —
spawns the task, drives `app.update()` in a wall-clock-bounded
loop with `std::thread::yield_now()` so the shared
`AsyncComputeTaskPool` gets a chance to schedule between polls.
- `winnable_seed_search_drops_in_flight_task_on_new_request` —
fires a winnable-only request, then before the task can complete
fires an explicit-seed request that bypasses the solver entirely.
Asserts the explicit seed wins, verifying the cancel-on-replace
contract.
Existing solver tests pass unchanged: explicit-seed paths skip the
new branch and run synchronously like before.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
funman300
d489e7a31b