[kvlonge]

Why

Keep compiler/planner-owned metadata and graphs native instead of constructing Python containers only to read them back.

Changes

• Retain header audits, column contracts and dynamic-family metadata; preserve custom constructors and opaque Python values.
• Retain planner graph/model/selection data through native consumers, with lazy public record projection.
• Remove obsolete private interfaces, document intentional Python edges, and lower crossing budgets.
• Add integration coverage and selectable DuckShop header-contract/shared-ancestor cases.
• Keep migration graph walks and model selection native; remove public orb bootstrap in favor of private project settings.

Verification

• make check-ci passed; affected Rust crates: 362 passed; initial focused Python: 49 passed; migration/planner regressions: 951 passed.
• Final CI passed: static/unit/integration, warehouse E2E, wheels, platform matrices and performance guards. Fixed the initial five planning failures caused by repeated scope projection; all five also pass locally.
• DuckShop: 293 models, zero errors/warnings; isolated DuckDB edge tests passed.
• Final same-machine dense/fresh 3,000-model comparisons: three alternating runs, no-cache/cache-write/warm/edit, identical reports/artifacts. Worst wall/CPU/RSS ratios: 1.025/1.047/1.017. Shared-ancestor planning: 8.759s base, 8.778s head medians, identical JSON.
• Independent review findings resolved; hygiene scan passed.

Verification is complete with no known blockers. External real-project validation is optional and is not a release or deployment requirement.