fmt: preserve ') VALUES (' seams and !=/<> operator spellings

[kyleconroy]

Two formatting fixes found by running sqlc fmt on a real project (sqlc-dev/cloud#604).

) VALUES ( seams broke apart

The boundary before a bare VALUES list was decided by comparing the source lines of the neighbouring nodes — the last column name and the first value. When both paren lists are multi-line those always sit on different lines, so an authored seam was forced apart:

INSERT INTO archives (
  id,
  project_id
) VALUES (          -- was reformatted to `)` + newline + `VALUES (`
  $1,
  $2
);

The seam now gets its own layout group in InsertStmt.Format, and AttachComments reads the author's actual choice out of the source: scanning back from the first value across ( and the VALUES keyword, the whitespace in front of the keyword says whether the author broke the line. Glued stays glued; VALUES on its own line stays on its own line. Source that doesn't scan cleanly (an extra paren, a comment against the keyword) falls back to the old heuristic.

!= and <> were rewritten to each other

PostgreSQL's grammar normalizes != to <> before the AST exists, so fmt printed <> for authors who wrote !=. The MySQL converter mapped opcode.NE to != unconditionally, rewriting in the other direction. Both format parsers now read the author's spelling back out of the source, the same way the SQLite converter already did:

• PostgreSQL gets a NewFormatParser that rewrites A_Expr names at the operator's recorded location (the A_Expr location points at the operator token). The compiler keeps using NewParser, which preserves the canonical <> the catalog knows.
• The dolphin format parser (already split off for identifier-case preservation) threads the source into the converter and checks the token in front of the right operand.

Testing

• Extended internal/endtoend/testdata/fmt/{postgresql,mysql} with a multi-line ) VALUES ( INSERT, a broken-seam INSERT, and mixed !=/<> queries; regenerated goldens. The new cases produce no diff hunks — fmt leaves them exactly as written.
• Full TestReplay/base corpus and TestFormat pass.
• Re-ran the fixed fmt on the cloud repo's queries: seams and spellings preserved, output still AST-equivalent, sqlc fmt idempotent on the result.

🤖 Generated with Claude Code

https://claude.ai/code/session_01T4UpYmwPdxzePtsRXoL1H8

---

Generated by Claude Code