[kvlonge]

Why

The benchmarks page repeated the same comparison across seven tables and long methodology sections, which made the headline hard to find.

Changes

• Lead with one table: SQLBuild default vs dbt v2 strict with parse cache (cold, unchanged, one-model and 1% edits, ratios), and one line for cold CPU and memory.
• Three bullets replace the reuse-off, every-rule and --dirty tables, including where dbt --dirty is faster.
• One compact all-variants table (wall time) links the run for CPU and memory.
• Shorter "What is compared", a ✓/✗ correctness matrix, the reuse section trimmed, and a short reproduce section pointing to the harness README for per-variant commands.
• Same numbers, all from run 39543189999. The #compiler-guards-in-ci anchor is kept. Docs skill pages regenerated.

Verification

• Every number checked against the run's summary.md.
• npm run build in website/ passes; the anchor resolves.
• scripts.check_docs_skill passes; hygiene scan clean.