[kvlonge]

Why

The benchmarks page showed 3k/5k numbers from SQLBuild 0.113.3 and no comparison with dbt. The manual benchmark workflow has now run on a GitHub-hosted runner with released packages, with strict analysis proven to have run.

Changes

• benchmarks.mdx now leads with SQLBuild 0.164.3 vs dbt 2.0.9 on the same 10,000-model DuckDB project, from run 39543189999 (4 CPUs, 15 GiB RAM, medians of 3 rounds):
  • SQLBuild default vs dbt v2 strict with parse cache: cold, unchanged (replay), one-model and 1% edits, with CPU, peak memory and ratio. Also SQLBuild with reuse and delta compile off.
  • dbt v2 --dirty, strict and not. It is faster on a one-model edit (1.6 s, 48 models); at 100 edits it falls back to a full parse.
  • dbt v2 default and remote-source strict rows as context.
  • SQLBuild with every rule, the correctness matrix, and what is and isn't compared.
  • How reuse works, then project, methodology, exact commands, reproduction with gh workflow run and locally, and the CI guards.
• The 0.113.3 tables are replaced.
• concepts/rules.mdx links to the new section; the docs skill pages are regenerated.

Verification

• Every number checked against the run's summary.md.
• The reuse section checked against cli/compile.mdx.
• npm run build in website/ passes and the page renders.
• scripts.check_docs_skill passes; hygiene scan clean.