Milestones 6 and 7: DDL and utility statements — parse corpus at 100%

[kyleconroy]

Implements milestones 6 (DDL part 1) and 7 (DDL part 2 + utility) from PLAN.md, driving the parse suite's todo list from 19,787 cases to zero. The gate for milestones 4–7 — 100% of the regress corpus with byte-identical trees, JSON, error messages, and cursor positions — is met; the scan and both split suites are also at zero.

Milestone 6

• CreateStmt: table elements, column/table constraints (CHECK, UNIQUE, PRIMARY KEY, EXCLUDE, FOREIGN KEY with key match/actions), LIKE clauses, identity/generated columns, partition specs and bounds, typed (OF type) and PARTITION OF tables
• CreateTableAsStmt including the AS EXECUTE form, with bounded backtracking for the CREATE TABLE t (…) element-list vs. CTAS column-list ambiguity (commits only at AS)
• AlterTableStmt over TABLE/INDEX/SEQUENCE/VIEW/MATERIALIZED VIEW/FOREIGN TABLE: the full alter_table_cmd set, ATTACH/DETACH PARTITION, AlterTableMoveAllStmt, identity column options, ALTER TYPE … ATTRIBUTE (composite types)
• IndexStmt, ViewStmt (recursive views via makeRecursiveViewSelect), CreateSeqStmt/AlterSeqStmt, CreateMatViewStmt/RefreshMatViewStmt, CreateSchemaStmt with nested schema elements

Milestone 7

• Functions/procedures with SQL routine bodies (BEGIN ATOMIC), ALTER FUNCTION, triggers and event triggers, rules, policies
• GRANT/REVOKE, role grants, ALTER DEFAULT PRIVILEGES
• Generic DROP/COMMENT/SECURITY LABEL object machinery (function_with_argtypes and friends)
• Unified ALTER dispatcher: RENAME / SET SCHEMA / OWNER TO / DEPENDS ON EXTENSION across ~30 object kinds, each arm restricted to the tails its object supports
• DefineStmt family (aggregates, operators, types, collations, text search), operator classes/families, extended statistics, domains, databases, extensions, FDWs/servers/foreign tables/user mappings, publications/subscriptions, tablespaces, conversions, transforms, procedural languages
• EXPLAIN/VACUUM/ANALYZE/CLUSTER/REINDEX/DO, transactions, SET/RESET/SHOW, LOCK, TRUNCATE, NOTIFY/LISTEN/UNLISTEN, and the rest of the utility set
• SplitWithParser and IsUtilityStmt now have real implementations (ported from pg_query_split.c / pg_query_is_utility_stmt.c) with unit tests

Byte-parity details the corpus forced out

• C's zero-valued enums surface as proto-shifted spellings: every AlterTableCmd/RenameStmt carries behavior: DROP_RESTRICT, RenameStmt.relationType defaults to OBJECT_ACCESS_METHOD, AlterPublicationStmt.action to AP_AddObjects
• NIL-vs-empty-list shapes: the (*) aggregate stores a null list element in DefineStmt.args; an empty BEGIN ATOMIC body is [NIL], not [[]]
• Support-function ereports carry their C call site's funcname (processCASbits, SplitColQualList, makeOrderedSetArgs, preprocess_pubobj_list, …); ConstraintAttributeSpec's location is always -1 (PG's YYLLOC_DEFAULT gives empty productions -1), so processCASbits errors report no cursor position
• base_yylex-merged *_LA tokens are never identifiers (found via nulls first in index_elem)
• makeRangeVarFromAnyName leaves inh false (C makeNode semantics), unlike makeRangeVar

Tooling and docs

• cmd/difftodo: debug aid printing input/want/got for a file's remaining parse todos
• PLAN.md updated with as-built notes for milestones 6–7

Testing

• go test ./... green; every graduated case is regression-locked (a future mismatch is a hard failure, never a new todo)
• go test -race over the parser and internal packages passes
• Remaining todo suites (deparse, fingerprint, normalize, normalize_utility) belong to milestones 8–10

🤖 Generated with Claude Code

https://claude.ai/code/session_01J19Kf5LpGUhi2vxFJjzQJE

---

Generated by Claude Code