[kvlonge]

Why

Large projects paid repeated planning costs that scaled with both model count and SQL-test count. Each selected SQL test rebuilt the same project-wide fixture metadata, complete authoritative fixtures still triggered required-column analysis, and broad Snowflake relation inspection used unnecessarily small serial query chunks.

Changes

• build and reuse one fixture-planning context per selected test set
• skip required-column analysis when an inferred fixture already supplies every authoritative column
• preserve per-test expected-type propagation without mutating shared metadata
• increase bounded Snowflake bulk column inspection chunks from 200 to 2,000 relations
• add deterministic work-count guards plus compile-to-plan integration coverage

Verification

• focused planner, scenario, required-column, and Snowflake tests: 114 passed
• fixture work-guard and integration tests: 3 passed after test-structure updates
• make check-ci
• independent read-only correctness review: no findings