[kvlonge]

Why

Semantic completion runs after model analysis. It recovers output types for failed models, adds B002 recovery notes, explains diagnostics and rejects unneeded sql_analysis false opt-outs. It still runs in Python and calls the Polyglot wheel to tokenize and parse failed models. This is the first part of moving the stage natively behind the native-preview engine.

Changes

• New native module sqlbuild-analysis::semantic_checks ports:
  • type recovery: projection spans from the native tokenizer, poisoned outputs spread through compiled lineage, and the "downstream output uses were not type-checked" notes
  • B002 recovery and the "downstream uses ... were not checked" notes
  • explanations: closest-column help, schema evidence, alias qualifiers from the native parser, operand types and comparison help
  • opt-out rejection (P009) and the code-owned help texts
• Python passes plain data from CompiledProject and applies the result. Shape publication and the binding-catalog revalidation stay Python inputs.
• Anything the native code cannot reproduce exactly falls back to the Python stage. That covers non-ASCII text, unsupported dialects, unreadable SQL and native failures. Each fallback is recorded as a deferral.
• Metadata checks (B300–B302, B102, wheel site Feat/bigquery adapter #7) still run in Python and are recorded as metadata_checks deferrals. A follow-up PR ports them.
• The python and native engines are unchanged.

Verification

• Rust unit tests; a parity property test against Python on generated failing projects in four dialects; failure-corpus cases; a DuckDB CLI E2E comparing all three engines
• make compiler-differential with CI flags passes: 281 projects identical, analysis coverage 56/56, wheel sites feat: add shared type normalization #8 and Feat/databricks adapter #9 at 0 under preview
• make compiler-differential-shipped, tests/e2e/scripts/compiler_differential, cargo test --workspace --all-features, clippy and fmt on 1.99.0, make check-ci, make test, hygiene scan