design: single source of truth for SynaptExtraction IL (schema ↔ validators)

[laynepenney]

Context

Sprint 31 shipped @synapt-dev/extract 0.1.0 with handwritten JSON Schema + handwritten validators in TypeScript and Python. Atlas's adversarial sweep on extract#8 found substantial drift between four sources:

• Schema too loose vs validators
• Validators too loose vs schema
• Implementation drift between Python jsonschema and TS ajv on format keywords

Sprint 31 patched the highest-impact drift but cannot eliminate the class — every new field has 4 places to write it (schema, TS validator, Python validator, hosted schema copy) and 6 pairs that can drift.

Goal

Single source of truth: schema, validators (TS + Python), and generated types all derived from one canonical source. Adding a field is one edit. Drift is impossible by construction.

Architecture options

1. JSON Schema as source → generate validators (data-driven, schema canonical). Semantic checks stay declared once, ported to both languages.
2. TypeBox as source → generate schema + TS types. Python validates against generated schema.
3. Pydantic v2 as source → generate schema and TS types via .model_json_schema(). Python primary.
4. Hybrid: schema canonical for structural; handwritten + ported semantic layer for cross-cutting.

Acceptance test

Atlas's fixture corpus from extract#8 review. Every fixture must produce identical verdicts across the hosted JSON Schema (third-party validator), TypeScript runtime, and Python runtime.

Scope

Sprint 32 design + implementation. Ship 0.2.0 from this work. 0.1.x stays as the handwritten beta. Migration notes in CHANGELOG.

Out of scope

• New IL fields (separate stories)
• Prompt-system refactor (independent)