feat(ddl): CREATE SEQUENCE / DROP SEQUENCE + nextval/currval/setval helpers

[productdevbook]

Summary

• First-class CREATE SEQUENCE / DROP SEQUENCE DDL on PG and MSSQL — a free-standing monotonic integer source independent of any particular table's lifecycle (auto-increment/IDENTITY columns are table-scoped; sequences are not). Useful for advisory IDs, batch numbers, cross-table counters.
• PG-only nextval('seq') / currval('seq') / setval('seq', n[, is_called]) runtime helpers wired into eb.ts and exposed as flat sumak exports.
• MySQL and SQLite refuse both surfaces via the new SEQUENCES / SEQUENCE_FNS feature flags; the runtime gate fires inside BasePrinter.printFunctionCall so the refusal lands at print time with a clear error rather than driver-level parse failure.

Surface (new)

• db.schema.createSequence(name [, schema]) → CreateSequenceBuilder with .ifNotExists() .dataType(...) .increment(...) .minValue(...) .noMinValue() .maxValue(...) .noMaxValue() .start(...) .cache(...) .cycle() .noCycle() .ownedBy(table, column) .ownedByNone().
• db.schema.dropSequence(name [, schema]) → DropSequenceBuilder with .ifExists() .cascade().
• Flat exports: nextval(seqName), currval(seqName), setval(seqName, value [, isCalled]). Sequence names are validated against a SQL-identifier regex; numeric value must be a finite integer.

Dialect divergence (handled at print time)

Feature	PG	MSSQL	MySQL	SQLite
CREATE SEQUENCE	yes	yes	refused	refused
IF NOT EXISTS	yes	refused (pointer at wrapper pattern)	—	—
OWNED BY …	yes	refused (no analogue)	—	—
DROP SEQUENCE … CASCADE	yes	refused (sequences aren't in the ref graph)	—	—
nextval/currval/setval	yes	refused (use NEXT VALUE FOR)	refused	refused

Test plan

• pnpm fmt && pnpm lint && pnpm typecheck && pnpm vitest run all green (2286 passed, 58 skipped — +38 from new file).
• New file test/ddl/sequences.test.ts with 39 tests covering builder DSL, dialect emit shapes, refusal paths, and a PGlite roundtrip (create → nextval × 2 → currval → setval → nextval → drop).
• gh pr checks <N> --watch green.

Tradeoffs

• MSSQL's NEXT VALUE FOR <seq> is intentionally not supported in this cut — it's a sequence-value expression (not a function call), so cleanly bridging it needs a dedicated AST node and printer pass. Callers who need MSSQL sequence access today can fall back to unsafeRawExpr("NEXT VALUE FOR " + quotedName) when the name is trusted.
• Sequence names land as SQL string literals in the emitted SQL (nextval('order_no_seq')), not bound — PG's plan cache treats them that way, and binding would force per-call replan. The builder rejects names that aren't bare SQL identifiers to keep injection out of the literal slot.

🤖 Generated with Claude Code