Allow arbitrary arguments order for Q3C, Q3CN, and Qwen3.5

[sayap]

This should fix the read file at offset/limit issue, where the tool definition has offset before limit, while the model sets limit before offset.

• I have read the contributing guidelines
• Self-reported review complexity:
  • Low
  • Medium
  • High

[sayap]

https://www.reddit.com/r/LocalLLaMA/comments/1ri3y89/my_last_only_beef_with_qwen35_35b_a3b/ - Qwen3.5 thread, discussed in #1338 (comment)

https://www.reddit.com/r/LocalLLaMA/comments/1r6h7g4/qwen3_coder_next_looping_and_opencode/ - Q3CN thread

[ubergarm]

Ahh I had trouble with various qwen35moe models looping when experimenting/testing with opencode as well.

I have not tried the work-around mentioned in the Q3CN thread ~/.opencode/plugins/read-limit.ts.

In a single test this seems like it possibly fixes the issue, but I'll have to try some more tests. Also didn't regression test other models.

---

wait, @sayap do i need to run llama-server --relax_arg true or something to use this? i did not do that in my test so might be moot.

[sayap]

Thanks for testing. I just checked the mainline autoparser branch as well, and I think it still lacks the parallel tool calls fix from #1300, and the arbitrary arguments order fix from this PR.

[sayap]

do i need to run llama-server --relax_arg true or something to use this?

form.relax_arg = true is set in common_chat_params_init_qwen3_coder_xml. No change to CLI