[kvlonge]

Why

SQL-test planning sent Python-built JSON to a native planner. Under preview the request now comes straight from the compiled objects.

Changes

• sqlbuild-analysis: public request/plan types; typed entry points for batches, chains and error messages. JSON entry points wrap them.
• Bindings plan_compiled_sql_tests, resolve_compiled_sql_test_chains read compiled objects; Python supplies adapter facts.
• Behind SQL_TEST_GLUE (preview): compile artifacts, sqb test planning (now using native error messages) and chains. Dead None fallback removed; JSON path kept for python/native.

Verification

• Parity test vs the JSON path, 5 adapters, two seeded variants; native gates count only binding results.
• Sabotage (temporary, removed): forcing the JSON path zeroes every native count; gates fail, parity holds.
• Rust typed-vs-JSON tests prove serde losslessness, not independent parity. Failure-corpus cases; preview cases in cursor-window and test --inspect E2E.
• Seams, cursor-window, inspect, helper-reference, fast_lineage, differential E2E; make check-ci, make test, hygiene. Earlier: cargo test, clippy/fmt 1.99.0, make compiler-differential (307), -shipped (296).

Test expectation changes

• Seams E2E: plan_native_sql_test_artifacts None -> artifacts, one plan_compiled_sql_tests call per batch.
• Thresholds are deterministic seeded counts. Failing variant adds helper redefinitions and unflattenable assertions, so more batches raise: planned 850->624, with errors 150->144, chains 350->398; new native-raised ≥102 covering compile_input, planner_input, sql_test_reference. Window "raised ≥40" is now a corpus floor (≥108 raised outcomes, either path).
• New clean variant (batch ordering): 30/30 batches answered; planned ≥1068, errors ≥192, chains ≥122, native-raised ≥45.
• The compiler reports unmocked references first (P013), so each project gets one appended unmocked assertion to reach sql_test_reference.