Milestone 5: DuckDB-native statement transformers

[kyleconroy]

Milestone 5 from PLAN.md: transformers for every remaining statement family, ported from the pinned DuckDB commit (8cbdaba6ac), plus the shape-test coverage for statements json_serialize_sql cannot see. Every Statement alternative now has a transformer — Parse no longer reports ErrUnsupported and the unsupported-statement registry is gone.

New statements (ast/misc.go + parser/transform_misc.go, transform_copy.go, transform_create_misc.go)

• SET/RESET — all scopes (LOCAL/SESSION/GLOBAL/VARIABLE), the SET SCHEMA and SET TIME ZONE special forms, and upstream's DEFAULT-to-RESET folding. PRAGMA — call form, assignment-to-SET desugaring, SQLite-compat pragmas (table_info), named parameters.
• CALL, USE, CHECKPOINT, transactions, VACUUM/ANALYZE, EXPORT/IMPORT DATABASE, PREPARE/EXECUTE/DEALLOCATE, EXPLAIN — with the preparable-statement check, positional-argument naming for EXECUTE, and EXPLAIN (options) handling.
• ATTACH/DETACH/CONNECT/DISCONNECT, COMMENT ON, INSTALL/FORCE INSTALL/LOAD/UPDATE EXTENSIONS, and the external-resource statements.
• COPY in all forms: table TO/FROM with the generic and PostgreSQL-style specialized option lists, format extraction from extensions, COPY (SELECT) TO, COPY FROM DATABASE (flagged form and the PRAGMA copy_database desugaring), and COPY TO as a CTE body via a new CopyQueryNode (a DuckDB 2.0 feature darkwing previously rejected).
• CREATE MACRO / SECRET / TRIGGER as CreateInfo variants: macro overloads and parameter defaults, secret TYPE/PROVIDER/SCOPE extraction with default naming, trigger timing/events/REFERENCING/FOR EACH/body.

Where upstream desugars a statement into another family (USE→SET, CHECKPOINT→CALL, IMPORT→PRAGMA, DEALLOCATE→DROP, ANALYZE→VACUUM), darkwing keeps a dedicated node per PLAN.md's AST list and notes the upstream shape — for sqlc's benefit.

Transformer-raised Parser Errors

The corpus's todo entries expected upstream's transformer rejects to clear as transformers land, so this PR also ports that long tail: recursive-CTE restrictions (COPY/DML bodies, LIMIT/ORDER BY, duplicate names, n-ary union splitting), star EXCLUDE/REPLACE/RENAME interactions (with upstream's prefix-matching column comparator), WITHIN GROUP ordered aggregates, window-definition rules (frame copying, override checks, ORDER BY ALL, window functions inside definitions, frame-bound direction), VALUES row lengths, empty selection lists, CUBE/ROLLUP and grouping-set limits, sample ranges, cast-then-subscript, FK actions and column counts, CHECK subqueries, generated-column restrictions, sequence-option validation (duplicates, NULLs, evaluated range checks), ALTER restrictions, TIMESTAMP/TIME modifiers, array bounds, pivot checks (constants, subqueries, data-extracted pivots in views/macros/with parameters), MERGE unreachable clauses, and escape-string NUL/UTF-8 validation.

One milestone-3 behavior is corrected: mixing named and positional parameters now accepts — upstream raises NotImplementedException, which the corpus oracle classifies as post-parse.

Conformance

• 139 corpus todo entries cleared (227 → 88). The accept/reject gate agrees with the oracle on every non-todo case.
• The 88 remaining todos are oracle rejects raised outside the parse pipeline — bind-time validation that happens to throw ParserException (setting values, function argument checks, PRIMARY KEY verification) and nested SQL parsing inside query()/nextval(). Their notes now say so, pointing at milestone 6's error-fidelity work.
• parser/misc_test.go adds the milestone-5 shape tests (the meyer-style snapshot substitute for non-SELECT statements); the serialize goldens gate stayed green throughout.
• Full dev loop green: go build, go vet, gofmt, go test -race ./....

🤖 Generated with Claude Code

https://claude.ai/code/session_01SDnHdNTsKPFE18AvkqT38C

---

Generated by Claude Code