[kvlonge]

Why

Native enrichment sent CTE fact recovery for contract models back to Python. Under preview those deferrals are now 0.

Changes

• Native port of _polyglot_cte_passthrough_facts and filtered non-null outputs (cte_facts.rs); adapter nullability rules passed to native.
• Defers: non-ASCII text, unknown dialect, types the normaliser defers, unshipped rules, unusual CASE/star/join/CTE shapes, panics. No Python deleted: the default engine still uses it.
• Test-only binding _oracle_cte_fact_recovery (existing _oracle_* pattern).
• Out of scope: 12 legacy_analysis, 1 input_enrichment deferrals (same on main).

Test expectation changes

• Model-analysis parity: "≥20 CTE recoveries" counted Python's reference run, so it is now python_cte_recoveries ≥40, a corpus floor (deterministic 40). Native proof: 0 input_enrichment deferrals (max 120 -> 0) and no mismatches. Native-model minimum also subtracts enrichment deferrals.
• New test_native_cte_fact_recovery.py: native vs Python recovery over 5,000 seeded queries, adapter profiles (duckdb, snowflake, bigquery, databricks, postgres, tsql, generic, None). The end-to-end test misses most differences (enrichment output only fills untyped columns), and no file compares the functions. Deterministic floors: 4,319 compared, recovered types 1,243, nullability 417, direct 859, non-null 38; 681 non-ASCII deferrals.
• The wheel parses IF/IFF as if_func, whose arguments Python never collects, so adapters' IF/IFF rules are unknown on both sides. A synthetic kind-keyed profile reaches conditional_result/first_arg.

Verification

• Sabotage (temporary, reverted): conditional_result and first_arg each fail the differential (2 and 3 mismatches). Forcing native recovery to defer fails the parity gate (68 <= 0).
• cargo test -p sqlbuild-analysis, clippy/fmt 1.99.0, parity files, make check-ci, make test, hygiene. Earlier: differential (319 identical), seam/differential E2E.

[kvlonge]

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