Maya refines the data model with LLM assistance

[danielnaab]

User Story:

As a form creator (Maya), in order to improve the extracted specs without editing JSON, I want to describe changes in natural language and have the system update the DataCollectionSpec accordingly

Preconditions:

• A FormProject with DataCollectionSpec exists (Slice 2+)
• Maya is authenticated
• Comparison UI available (Slice 4)

Acceptance Criteria:

• Maya can describe a change in natural language (e.g., "add a phone number field to the contact section")
• LLM interprets the request and modifies the DataCollectionSpec via tool calls
• Changes are presented as a proposal using the comparison UI from Slice 4
• Maya can review the semantic diff before accepting
• Maya can accept or reject each proposed change
• Accepted changes update the spec; rejected changes are discarded
• Multiple refinement rounds supported in a single session

Success Metrics:

• Percentage of natural language requests correctly interpreted on first attempt
• Average refinement rounds needed to achieve desired spec state

Notes:

• LLM uses tool calls: add_field, remove_field, update_field, move_field, set_condition, etc.
• Tool calls produce auditable changes (each maps to a spec modification)
• Reuses comparison protocol from Slice 4 for review
• This is the authoring agent pattern from the design spec

Definition of Done:

• Acceptance criteria met
• Threat model updated -- any new trust boundaries, data flows, or attack surfaces are reflected in catalog/architecture/threat-model.md
• Technical documentation updated -- architecture docs and decisions are current
• Tool call interface defined and implemented
• Tests pass
• Type checking passes
• CI pipeline green
• Deployed and demoable