The local LLM is incapable of generating a valid prototype

### Describe the bug

When using the local LLM `ai/gpt-oss`, the LLM cannot generate a valid prototype and fails every time.

- It cannot correctly generate the `next_question_value` that points to the number of the next question.
- It uses the wrong `answer_type`, e.g., "text" for an address question
- It adds extraneous properties, e.g., adds options to questions that don't need them.
- It adds an empty "submit" question at the end.

### Steps to reproduce

Run the Docker Compose script with the local LLM `ai/gpt-oss`.

### How should the bug be fixed?

Options include:

- Try out different models.
- Improve the prompt & schema to coax the model. But need to be careful about tuning them for a local LLM to the detriment of the Azure LLM.
- Split the generation step into multiple steps (e.g., generate the questions, then put them in order). This is a larger piece of work but likely necessary anyway to generate larger prototypes

### Device, OS and browser used

_No response_

### Additional context (if applicable)

Follow-on from #148 and #152.

### Screenshots (if applicable)

_No response_

Reproduce the failure with Docker Compose and the local ai/gpt-oss model, then inspect the prompt and schema used for prototype generation. Compare the output against the required next_question_value, answer_type, question properties, and terminal submit behavior; done means valid prototypes are produced without regressing Azure LLM behavior.

Written by the indexing model from the issue text.