Closes Quat investigation #1. Today some Klondike deals are
unwinnable from the start and the player has no signal that the
deal they were given is solvable. A new Settings → Gameplay toggle
"Winnable deals only" (default off) makes the engine retry seeds
at deal-time until the solver returns Winnable, up to a cap.
Solver
solitaire_core::solver is a hand-rolled iterative-DFS solver with
memoisation on a 64-bit canonical state hash. Move enumeration is
priority-ordered: foundation moves first (zero choice when an Ace
or rank-up exists), inter-tableau moves second, waste-to-tableau
third, stock-draw last. The draw is skipped when the cycle counter
shows we've recirculated the entire stock without progress —
Klondike's deterministic stock cycle means further draws can't
unlock anything new.
Two budget knobs (move_budget = 100k, state_budget = 200k by
default) cap pathological cases at Inconclusive; the caller treats
Inconclusive as "winnable" so the player isn't penalised for the
solver giving up. Median solve time is 2 ms; pathological
inconclusives top out near 120 ms.
Switched from recursive to iterative DFS after a real-deal solve
overflowed Rust's default 8 MB thread stack. Behaviour identical;
the change is invisible to callers.
Pure logic — solitaire_core has no Bevy or I/O. Same input always
yields the same SolverResult.
Settings
Settings.winnable_deals_only is a #[serde(default)] bool; legacy
files load to false. SOLVER_DEAL_RETRY_CAP = 50 caps the retry
loop. The Settings → Gameplay toggle reads as "Winnable deals only"
with a "(may take a moment when on)" caption.
Engine integration
handle_new_game's seed-selection path now branches on the toggle.
When on AND mode is Classic AND no specific seed was requested
(daily challenges, replays, and explicit-seed requests bypass the
solver), choose_winnable_seed walks seed N, N+1, N+2, … calling
try_solve until it finds Winnable or Inconclusive. If the cap is
hit without a verdict, the latest tried seed is used so the player
always gets a deal rather than spinning forever.
19 new tests (11 solver, 3 settings, 5 engine including the
choose_winnable_seed unit). Two ignored bench/scan helpers
(solver_bench, find_unwinnable) for ad-hoc profiling.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
funman300
8a5fa8751c