Add builder-driven full extraction flow with embeddings

[laynepenney]

Context

The current builder work focuses on coupled prompt + Stage 1 response schema generation. That is useful, but we also need a first-class end-to-end path that runs the full extraction packet, including embeddings, through the builder.

This should cover both model calls:

1. extraction call: build prompt + response format, call the extraction model, parse Stage 1 JSON
2. embedding call: build embedding inputs, call the embedding model, attach embedding metadata

Then the builder should finalize the output with the full producer/source context rather than requiring callers to manually stitch these pieces together.

Scope

Add an API that can execute a complete extraction using builder configuration:

• extraction model/provider/options
• embedding model/provider/options
• source metadata such as source_id, source_type, user_id, and kind
• full produced_by metadata for the extraction model
• embedding records with model, input, dimensions, space, and computed_at
• capability/profile selection shared by prompt, schema, finalization, and verification

The builder should still support the current dry-build mode that returns prompt/schema/response format without making network calls.

Design Questions

• Should this live directly on ExtractionBuilder as execute() / run(), or as a separate runner that consumes builder.build()?
• Should the library ship provider-agnostic interfaces plus OpenAI helpers, or OpenAI-only helpers first?
• Should embeddings be configured as a single embedding of the source text by default, with optional summary/entities embeddings later?
• How should we represent auxiliary producer metadata for the embedding model versus the final extraction-producing model?

Acceptance Criteria

• TypeScript and Python both expose a builder-driven E2E API.
• One fixture demonstrates a real extraction API call plus a real embedding API call, not fabricated payloads.
• Finalized extraction validates against SynaptExtraction.
• The output includes structured produced_by for the extraction model.
• The output includes embeddings[] records with model URI, dimensions, input, vector, and computed timestamp.
• Tests cover prompt/schema parity, finalization, validation, and the embedding metadata path without requiring live API calls by default.
• Docs show both dry-build and execute-with-embeddings flows.