Qwen 3.5 template tweak stops needless prompt reprocessing

This is a practical community-level patch for a real latency pain point, even if it is not an upstream architectural fix.

• –The root cause is template-level prompt mutation, not just model size or hardware limits.
• –The workaround is narrowly scoped to thinking-disabled flows, so it avoids changing normal reasoning-mode behavior.
• –Keeping empty think tags is a low-context-cost tradeoff compared with repeatedly invalidating cache reuse.
• –The author reports good coding-session behavior but no long-context benchmarks yet, so production users should validate on their own workloads.