Summary

• New rocky compile --with-seed flag: runs data/seed.sql against in-memory DuckDB, reads information_schema, and feeds the result as source_schemas into the typechecker. Leaf .sql models pick up real types instead of Unknown.
• Opt-in (preserves the wave-1 fast-pure compile path), feature-gated behind duckdb.
• Reuses the existing sync DuckDbConnector + one information_schema.columns round-trip — no new dependencies, no async surface.

The reframe behind this PR

Orientation revealed Arc 7 wave 2's original framing was wrong. .sql and .rocky models go through the same compute_model_typecheck path (rocky-compiler/src/typecheck.rs:426) — there is no asymmetry to fix. The actual gap is source_schemas: HashMap<String, Vec<TypedColumn>> is initialized empty in every CLI command (grep "source_schemas: std::collections::HashMap::new()"). Leaf models reading from raw sources start Unknown and propagate it through the entire DAG. Wave 2's job is to fill that map; this PR fills it from one source of truth scoped to the playground audience for 1.0 launch.

The active-priority memory has been updated to reflect this reframe so the next session doesn't re-litigate.

Trust outcome made concrete

Run on the playground default POC:

WITHOUT --with-seed → raw_orders.incrementality_hint = {
  confidence: "medium",
  signals: ["column name 'order_date' ends with '_date' (timestamp pattern)"]
}

WITH --with-seed → raw_orders.incrementality_hint = {
  confidence: "high",
  signals: [
    "column name 'order_date' ends with '_date' (timestamp pattern)",
    "column type is temporal (Date), suitable as a watermark"   ← only possible with typed source
  ]
}

The type-derived signal cascades into incrementality recommendations, drift detection, P002 blast-radius accuracy, and any future feature that consumes TypedColumn.

What's deferred to wave 3

• Real-warehouse audience: cached DESCRIBE TABLE from rocky discover → compile. Different audience (production users), different cache infra (new write path + new persistent format + invalidation story). Multi-PR. The plug-in point — source_schemas — is the same; both options just fill it from different inputs.
• --with-seed auto-detection when data/seed.sql exists adjacent to models dir.
• Multi-file seeds, alternative seed paths, non-DuckDB seed backends.

Test plan

• 4 new integration tests in commands::compile::tests (all #[cfg(feature = "duckdb")]):
  • with_seed_populates_source_schemas_for_leaf_model — happy path
  • with_seed_bails_when_seed_file_missing — opt-in implies the user knows the path; hard fail with clear message
  • with_seed_bails_when_seed_sql_invalid — surfaces underlying DuckDB error
  • with_seed_resolves_unknown_types_to_concrete — asserts BIGINT → RockyType::Int64 and VARCHAR → RockyType::String round-trip through the loader
• All 10 existing compile tests still pass with the new signature
• cargo test --workspace — 0 failures
• cargo fmt --check + cargo clippy --workspace --all-targets -- -D warnings
• pytest in integrations/dagster/ — 307/307 pass; fixtures unchanged (the playground default fixture compile pipeline doesn't use --with-seed, so no fixture-side regen needed)
• just codegen — no schema or generated-binding drift (the new flag is CLI-only, doesn't change any output struct shape)
• End-to-end smoke test on examples/playground/pocs/00-foundations/00-playground-default/ — incrementality confidence diff shown above