docs(adr): source-modeling discipline — Track G G0

[ronsse]

Summary

• Adds docs/design/adr-source-modeling-discipline.md (~235 lines) codifying the project's stance on graph granularity decisions, with columns as the canonical worked example.
• Default rule: columns are properties on the parent table node, not standalone nodes. properties.columns (structured) + properties.column_names (flat denormalized list for exact-match search via existing in filter).
• Enforcement: warn + opt-in flag (allow_structural_leaf: bool = False) on EntityRule and EntityDraft. Soft validator emits structlog WARNING on entity_type in {"Column", "column", "TableColumn", "table_column"} blocklist unless flag set. Does NOT reject mutations.
• Regulated-column exception: operator sets allow_structural_leaf=True AND node_role="structural" for the rare case where column-level entities earn their own nodes (PII / governance lifecycle).
• Column-level lineage: no name-match. Stable column identifiers required; cross-table same-name flow is forbidden. v1 column-level lineage stays as table-level edges with optional edge.properties.column_pairs annotations.

Test plan

• ADR reads cleanly alongside docs/agent-guide/modeling-guide.md (Schema explosion anti-pattern) and docs/agent-guide/source-modeling-cookbook.md (existing column guidance)
• Cross-references to adr-canonical-graph-layer.md §4 + §8.2 resolve (the deferral of nested-path DSL operators)
• Blocklist constant name (ENTITY_TYPE_ANTI_PATTERNS) matches sibling Unit G1's implementation
• Validator helper name (validate_entity_type_not_anti_pattern) matches sibling Unit G1's implementation
• Sibling Unit G2's cookbook recipe is consistent with §2.2 (denormalization rationale)

🤖 Generated with Claude Code