extract or upstream structured JSON model-call utilities

[autholykos]

Context

Pituitary already implements a useful structured model-call pattern in its analysis provider:

• serialize a bounded prompt payload;
• call Stroma chat;
• normalize JSON response text;
• decode into a typed target;
• classify provider/schema errors.

RLM-style products need this on every worker/subcall.

Proposed Capability

Either:

1. move the generic JSON-call helper down into Stroma and adopt it in Pituitary, or
2. extract Pituitary's generic JSON-call code into a public product-neutral package.

Acceptance Criteria

• The helper is not tied to specs/docs.
• Callers can decode into typed structs.
• Invalid/missing JSON returns a structured schema-mismatch error.
• Pituitary's existing provider-backed adjudication tests continue to pass through the shared helper.