[Feature] Replace --is-reasoning-model with unified --reasoning flag in regenerate_train_data.py

[msober]

Checklist

• 1. If the issue you raised is not a feature but a question, please raise a discussion at https://github.com/sgl-project/SpecForge/discussions/new/choose Otherwise, it will be closed.
• 2. Please use English, otherwise it will be closed.

Motivation

Problem

The current --is-reasoning-model flag in regenerate_train_data.py is a binary switch that only controls whether to save reasoning_content from responses. It doesn't support a common use case: using a reasoning model (e.g., Qwen3.5) with thinking explicitly disabled on the server side.

For example, SGLang supports disabling thinking via:

extra_body={"chat_template_kwargs": {"enable_thinking": False}}

But there's no way to pass this through the current CLI interface.

Proposal

Replace --is-reasoning-model (boolean) with --reasoning (choice) that supports three modes:

• none (default): Standard model, no special handling
• save: Save reasoning_content from responses (current --is-reasoning-model behavior)
• disable: Explicitly disable thinking on the server side via extra_body, and do not save reasoning_content

This eliminates ambiguous flag combinations and covers the missing use case in a single parameter.

Related resources

No response