Pipeline Data Flow

This page is a compatibility note, not the primary architecture spec.

The current canonical overview lives in:

• docs/reference/data-architecture.md

Legacy training details that still matter:

• raw CAN rows are normalized, parsed, and cached before graph materialization
• graph materialization receives an explicit segment config derived from representation_cfg at the pipeline boundary
• the runtime loader still uses budget-aware batching for variable-size graphs

What changed:

• representation config is now the primary user-facing surface
• raw storage, materialized views, and discovery/hypothesis data are split
• snapshot, temporal, multi-scale, sequence, and entity are explicit representations