Harden reasoning-model output cleanup

[xdCloudy]

Summary

Improve cleanup of reasoning-model output before Mindcraft uses it as a conversation response or memory summary.

Some reasoning-capable models can expose internal reasoning delimiters such as </think> or emit the configured *** stop sequence literally. Mindcraft already strips </think>, but the current implementation only takes the first segment after the marker.

Changes

• Add a shared _cleanReasoningOutput() helper
• Keep the content after the final </think> marker
• Strip trailing literal *** stop-sequence leakage
• Reuse the same cleanup for conversation responses and memory summaries

Why

During testing with an OpenAI-compatible local reasoning model, responses occasionally contained:

<thinking output>
</think>

!command(...)

[Sweaterdog]

quit spamming our repo with slop

[Sweaterdog]

whoopsies

[Sweaterdog]

Succinctly does what is required to fix it for all model endpoints.