Emit OpenAI strict-mode function schemas (`strict: true` + `additionalProperties: false`) from the export adapter

[dgenio]

Summary

Add an opt-in strict=True path to chainweaver/export/openai.py so exported function payloads satisfy OpenAI's structured-outputs requirements: strict: true on the function plus additionalProperties: false on every object schema (recursively), with all properties marked required as the spec demands.

Why this matters

OpenAI's structured-outputs / strict function calling guarantees the model returns arguments matching the schema, which is exactly the determinism story ChainWeaver sells. Today _build_payload emits a plain function schema (no strict, no additionalProperties: false), so users who paste the export into a strict tool definition get a runtime rejection. Closing this makes the export adapter a drop-in for the most reliability-conscious OpenAI users.

Current evidence

• chainweaver/export/openai.py _build_payload produces {"type": "function", "function": {name, description, parameters}} and only schema.pop("title", None) — no strict flag and no additionalProperties: false injection.
• chainweaver/export/_schema.py (model_input_schema_json) is the shared schema emitter where recursive transformation would live.
• examples/export_openai_anthropic.py and docs/ (## Export adapters in README) demonstrate the adapter, so a strict variant has an obvious documentation home.

External context

OpenAI's structured-outputs guide requires additionalProperties: false on every object and all keys listed in required when strict: true; passing a schema that violates this is rejected at request time.

Proposed implementation

1. Add strict: bool = False to flow_to_openai_function/tool_to_openai_function; thread into _build_payload.
2. Implement a pure recursive schema transformer (in _schema.py) that, on every object node, sets additionalProperties: false and promotes all properties to required, handling $defs/$ref, anyOf/oneOf, and arrays.
3. Document the constraint that optional fields must use nullable-with-default patterns under strict mode (OpenAI requirement) and surface a clear ExportError when a schema cannot be made strict-compatible.
4. Mirror any reusable transform for the Anthropic adapter if applicable (note relation, do not over-scope).

AI-agent execution notes

Inspect chainweaver/export/openai.py, chainweaver/export/_schema.py, chainweaver/export/anthropic.py, existing export tests under tests/. Keep the transform pure and deterministic (no network). Edge cases: nested models via $ref, lists of models, enums, fields with existing additionalProperties. Do not change the default (non-strict) output shape.

Acceptance criteria

• flow_to_openai_function(flow, strict=True) produces strict: true and additionalProperties: false on every object including nested $defs.
• A round-trip test validates the payload against OpenAI's documented strict constraints (structural assertions, offline).
• Default behavior is byte-for-byte unchanged.

Test plan

Unit tests for flat/nested/array/$ref/enum schemas under strict mode; assertion that every object node has additionalProperties: false; regression test for the non-strict default; example script update run in CI.

Documentation plan

README ## Export adapters section, examples/export_openai_anthropic.py, a cookbook note, CHANGELOG.

Migration and compatibility notes

Not expected to require migration (opt-in flag; default unchanged).

Risks and tradeoffs

Strict mode forbids genuinely-optional fields, which can surprise users with optional tool inputs — document the nullable-with-default pattern and fail loudly rather than silently producing an invalid schema.

Suggested labels

integrations, ecosystem, product