[kvlonge]

Why

Profiling a generated 10,000-model project showed compile work that does not change output: every analysis-cache hit normalised its SQL eagerly although only a few readers need it, and SQL-test planning re-tokenised the same model SQL for every test and comparison body (about 7% of instructions on a 1,000-model cold compile).

Changes

• Analysis-cache hits normalise their cleaned SQL only when read: batched before binding revalidation and case-sensitive publication, and on demand for Python readers (outcome_rows, binding-diagnostic locations) through a new resolved_cleaned_sql.
• defined_cte_keys skips tokenising text without the keyword with and returns a plain list (every caller treated "unscannable" as empty).
• SQL-test planning memoises each model query's reserved-CTE check for the run.
• New E2E test: a cached model's binding diagnostics match the cold compile, before and after a reference placeholder.

Verification

• Alternating A/B runs on the 10k benchmark project: cold 27.3 → 26.7 s; one-model edit 21.4 → 21.0 s.
• Identical compiled output and JSON against main for cold and edit compiles on a 1,000-model generated project and a private real project (compile only).
• compiler-differential with stage captures and coverage gates: 368 projects as expected.
• cargo test -p sqlbuild-analysis, clippy and fmt (stable); focused pytest on SQL-test planning, cache invalidation, binding catalog, native model analysis and the new test; make test; make check-ci.
• The new E2E fails when the deferred normalisation is removed.