feat: add premise-driven schema selection

[rlippmann]

Summary

Adds premise-driven schema selection to the generic refund_intake examples in Python and TypeScript while preserving the existing policy-driven behavior.

The examples now show two host-owned schema-selection paths:

• use refund_intake / use technical_support selects a schema from policy state.
• Saved order facts in premise select a schema through a deterministic intake-context rule.

The same ambiguous request, I need help with order A-100., can now select different schemas based on authoritative state rather than model output or user wording.

Why

Schema selection is a natural fit for premise because factual context can change how the same request should be interpreted.

This strengthens premise coverage outside prompt construction and shows:

authoritative premise -> host-owned intake context -> selected schema

without adding provider, live-model, or directive-drafter complexity.

Example metadata

• Enforcement point: schema selection
• Domain: customer order/support intake
• Technology/runtime: generic Python and TypeScript
• Example type: minimal example
• Does this use directive drafter? should normally be no: no
• Does this derive state from model output? must be no for primary examples: no

Checklist

• Repository scaffolding / governance change
• One primary enforcement point
• Explicit authoritative state
• No model-derived state mutation
• Adversarial stub or equivalent test
• Observable runtime behavior change
• Domain vocabulary is natural
• Framework is secondary to the enforcement point
• Docs updated
• Tests/smoke checks added where practical

Notes

Validation performed:

• ./scripts/validate_python.sh — passed, 178 passed, 6 skipped
• ./scripts/validate_typescript_fast.sh — passed
• npx --yes markdownlint-cli2 python/examples/schema_selection/refund_intake/README.md python/examples/schema_selection/README.md typescript/examples/schema_selection/refund_intake/README.md typescript/examples/schema_selection/README.md — passed, 0 error(s)

Ollama, LiteLLM, Vercel AI SDK, live-model paths, and directive-drafter behavior were intentionally left unchanged.