[kvlonge]

Why

Reduce cold compiler overhead while preserving normal validation, isolated custom rules, cache identities and artifact publication gates.

Changes

• Reuse conservative borrowed-tree binding and annotation-independent type evidence; retain normal validation for unsupported shapes and warning-producing joins.
• Stage bounded cold-project artifacts alongside selected rules, then publish at the normal successful write phase; preserve unchanged files and stale-file cleanup.
• Use one native analysis worker for SQL-heavy batches (10k+ projections, 48MiB+ prepared SQL) to bound compiler/custom-host memory; retain parallel scheduling for lighter batches.
• Avoid eager nullability fallback, unnecessary output chunking and unused custom-rule fingerprints when the rules cache is disabled.
• Add native differential and real CLI/DuckDB coverage for diagnostics, cache modes and publication failures.

Verification

• 163 focused Python tests passed; native query-analysis tests passed.
• uv sync --all-extras and make check-ci passed.
• The full cold 10k guard passed locally inside a 4GiB cgroup with zero cache hits and its expected fingerprint (107.09s). Its ceiling is 145s for memory-bounded scheduling; CI remains the final gate.
• Independent read-only review and bounded warning-preservation follow-up completed.