[kvlonge]

Why

Compile benchmarks at 10,000 models used a uniform synthetic project. DuckShop 10K grows the hand-written DuckShop project into a realistic large project, so scale guards measure the construct mix and graph shape of a real codebase.

Changes

• Add benchmarks/duckshop-10k/: a deterministic, versioned generator that writes a ~10,282-model Snowflake SQL project (about 60/19/21 staging/intermediate/marts) from DuckShop's patterns, with audits, sampled unit tests, scenarios, functions and a custom rule. Only the generator, its tests, a README and checksum.txt are committed; the output is never committed.
• --verify / --check compare the output with the checksum, a hash of a sorted per-file manifest. Output changes need a generator version bump.
• Add a cold-compile E2E test that generates the project, verifies the checksum, and asserts a clean compile within 60 s and 3 GiB.
• make test-e2e-dense-compile-performance runs the generator tests and that test in the 10K cell only, after the dense guards, so they never share the job's memory limit.
• Generator changes now turn on the performance CI scope.

Verification

• 27 generator tests, including a two-build determinism check under different hash seeds.
• A cold compile of the generated project: 0 errors, 0 warnings, about 16 s and 1 GiB locally.
• make check-ci; make -n for the 10K and 1K cells of the dense target.
• The hygiene scan on the repository and on the generated output.