Tiny World Auto Batching

Auto should not call the model on every placement.

Current intended model:

• Capture the current home board as sparse 8x8 JSON using snapshotCells().
• Ask the model for a ranked batch of candidate tile actions matching AUTO_SUGGESTIONS_SCHEMA.
• Spend autoSuggestionQueue locally for multiple Auto placements.
• Refresh after AUTO_REFRESH_EVERY Auto placements or when the queue runs out.
• Auto still places on the clicked cell. The model suggests tile/action definitions, not coordinates.

Implementation guardrails:

• Keep provider/model/key state shared with the Generate modal via AI_LS.
• Use normalizeAutoAction before applying model output.
• Use adaptAutoSuggestionToCell(action, x, z) so same-kind clicks increase local intensity.
• If no API key exists, open the Generate modal with the Auto key message.
• Do not add a backend or new dependency.

Validation:

• Stub generateAutoSuggestions in browser eval and verify two Auto suggestions only call the stub once.
• Verify no network call is needed while spending cached suggestions.