[kvlonge]

Why

Writing the compile cache made cold compiles 14-18% slower than --no-cache, and fresh compile was ~1.22x of 0.119.1. Profiling on 0.126.2 showed a whole-SQL identifier regex in binding-query preparation, dataclass construction just to serialise cache rows, full model SQL sent to native SQL-test chain resolution, and per-model scans in metadata validation.

Changes

• Qualified-reference scan searches each reference name directly (old regex kept as fallback for quoted, commented or non-ASCII SQL).
• Cache writes serialise compact lineage rows directly and compute each output signature once; stored payloads byte-identical, so no cache version change.
• SQL-test chain resolution sends names and dependencies only.
• Metadata validation serialises each select once and indexes argument column types.
• New guard: cold cache-writing compile vs --no-cache (CPU <= 1.10, wall <= 1.15, same fingerprint). Cache job warm/edit-to-cold ratios raised to 0.70 because cold got faster while edit and warm were unchanged (to be tightened with the warm-path work).

Verification

• Cache overhead at 3k: +18.3% -> +6.5% wall; dense 5k cold 42.5 s -> 37.2 s; fresh 3000 9.70 s -> 9.27 s (1.175x of 0.119.1, so fresh budgets stay as they are).
• Fingerprints, compiled artifacts, diagnostics and cache payloads identical to main.
• make test, make check-ci, cache and cold-compile performance targets under 4 GiB; one independent review.