[kvlonge]

Why

The benchmarks page needs a reproducible SQLBuild vs dbt comparison. Our shared dev machine is too loaded for fair wall times, so the benchmark runs on a GitHub-hosted runner from released PyPI packages.

Changes

• benchmarks/dbt-compare/: write_dbt_dense.py writes the CI dense project for SQLBuild (DuckDB), a rules-off copy and an equivalent dbt project (same SQL, names, folders, DAG; Jinja macros, unit_tests, data_tests, UDF functions). edit_models.py edits N model bodies. bench.sh times cold, warm, one-model edit and 100-model edit runs per variant over interleaved rounds. summarize.py prints median wall, CPU and peak RSS per cell. correctness.sh runs every tool on clean, missing-ref and missing-column projects. README covers method and what isn't equivalent.
• Variants: sqb_rules, sqb_norules, sqb_rules_full (reuse and delta off), dbt v2 parse, compile, compile --static-analysis strict, dbt Core parse/compile as reference.
• .github/workflows/benchmark-dbt.yml: workflow_dispatch only, on ubuntu-24.04 like the dense compile jobs, 300-minute timeout. Inputs pin sqlbuild (default latest), dbt 2.0.9, dbt-core 1.12.6, dbt-duckdb 1.11.0, models, rounds, variants. It uploads results plus summary.md (runner CPU/RAM, versions, correctness and benchmark tables) and writes the summary to the job page.

No product code changes, no raw results.

Verification

• actionlint on all workflows; release gate check passes.
• ruff format --check and ruff check on benchmarks/.
• Smoke run of bench.sh (all eight variants, 100 models, one round) and correctness.sh against the released packages: all runs exit 0 on the clean project. SQLBuild reports reuse and delta compiles where expected. Both seeded errors fail in SQLBuild; only the missing ref fails in dbt default compile.
• Hygiene scan clean.