[kvlonge]

Why

SQLBuild needs a realistic, dense Snowflake project for exercising the compiler, Rules, unit tests and scenarios, and for measuring Snowflake-to-DuckDB translation during local scenario replay. The existing example projects are too small to show real-world construct mixes.

Changes

• Add benchmarks/duckshop/: 252 Snowflake SQL models (staging, intermediate, marts, snapshots), 5 seeds, SQL and Python functions, macros, enums, constants, generic and singular audits, and reviewed rule exceptions for a fictional duck-themed home goods shop.
• Add 266 unit tests and 8 scenarios, with captured local snapshots.
• Add deterministic self-managed loaders for 76 managed sources (about 60,000 synthetic rows, scaled by the data_scale var), offline loader tests, a construct census script, and cleanup SQL.
• Add RUNBOOK.md for loading, building, testing and capturing on Snowflake, and a README with project status, expected divergences and known SQLBuild limitations.
• Add an E2E test that compiles the project with the SQLBuild under test and fails on any error or warning.

Verification

• sqb compile: 0 errors, 0 warnings; sqb format --check clean; ruff clean; 269 offline loader tests pass.
• On Snowflake with SQLBuild 0.175.0: sqb load, sqb build, sqb test (266/266) and sqb scenario test (8/8) pass, and sqb scenario capture succeeds.
• sqb scenario test --local --strict fails on 8 of 8, as documented in the README: 5 translation failures, and 3 snapshots reported stale by a SQLBuild fingerprint issue.
• make check-ci and CI cover the new E2E compile test.