[kvlonge]

Why

Under preview, SQL tests were still assembled in Python. This builds their facts natively and hands back only what it cannot match exactly.

Changes

• sqlbuild-analysis: native assembly of test names, target models, scope deps (incl. the macro call scan), case fingerprints (with Decimal.normalize) and P013 mock-reads-referencing-helper diagnostics with locations.
• Binding assemble_compiled_sql_tests reads compile inputs directly; SQL_TEST_GLUE assembles natively and reports diagnostics in Python's order.
• Per-test deferrals: macro mocks, non-ASCII text on the P013 path, decimals the context would round, failing reference scans. Macro mocks defer on purpose; M3 close-out Ports A routes them through the bridge. The differential records only macro_mocks (29).
• Kept: _assemble_compiled_sql_test, still used by the python and native engines.

Verification

• Assembly parity test (seeded model, direct, parameterized and mock-reads-helper tests) vs Python, plus Rust unit cases incl. a Python-computed fingerprint; P013 failure-corpus case with location.
• Sabotage: deferring every test fails the parity gate (0 native) and seams E2E. Reverted. The gate counts only assemblies compile consumes.
• Compile seam/deferral/lineage E2E plus test E2E (186 passed); make compiler-differential CI flags (321 identical); cargo test --workspace, clippy/fmt 1.99.0; make check-ci; make test (11940 passed); hygiene.

Test expectation changes

• Seams E2E: assemble_native_sql_tests expected None, now native assemblies (none deferred): the seam answers.
• Assembly parity counts are seeded and deterministic, so they are asserted exactly (was floors 150/13/80/3): native assembled 181, with P013 17, case fingerprints 96; deferrals capped exactly: macro_mocks 4, non_ascii_text 11, decimal_context 5. Any other deferral kind fails.
• Sabotage re-run on the exact gate: all-defer fails it (0 native, 201 deferred), parity intact.

[kvlonge]

Landed together in #1075 to cut merge-queue CI runs; review history stays here.