refactor(core): make KlondikeInstruction the move currency

Remove the (from, to, count) tuple as an internal move-passing wrapper.
Game logic now stays in KlondikeInstruction space end to end:

- Add GameState::apply_instruction, the native apply path. move_cards
  becomes a thin pile-coordinate adapter that converts to an instruction
  and delegates, so move bookkeeping (validation, score/recycle history,
  undo snapshot) lives in one place instead of being duplicated.
- next_auto_complete_move matches DstFoundation directly instead of
  projecting every candidate to pile coordinates.
- proptests and the storage round-trip test apply instructions directly
  rather than round-tripping instruction -> tuple -> move_cards.

The single instruction -> pile decode is renamed instruction_to_highlight
-> instruction_to_piles and kept in core: decoding a tableau run length
needs upstream pile-stack types core does not re-export, so relocating it
would duplicate the logic across engine and wasm. The two rendering edges
(engine hint highlight, wasm debug move list) call this one decoder; the
engine's hint_piles is a thin delegation to it.

Also includes the CardEntityIndex render-side index and a SelectionPlugin
init_resource fix so update_selection_highlight no longer panics in test
harnesses that omit CardPlugin.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

solitaire_wasm/src/lib.rs

@@ -357,6 +357,7 @@ fn legal_moves_for_game(game: &GameState) -> Vec<DebugMove> {
     let mut moves: Vec<DebugMove> = game
         .possible_instructions()
         .into_iter()
+        .filter_map(|instruction| game.instruction_to_piles(instruction))
         .map(|(from, to, count)| DebugMove::Move {
             from: pile_name(from),
             to: pile_name(to),