Fix sqlc fmt corruption in PostgreSQL and MySQL formatters

[kyleconroy]

Running sqlc fmt across all 800 PostgreSQL/MySQL endtoend configs and verifying each result (regenerated code compared with embedded SQL stripped, plus a parse-tree equivalence check for MySQL) surfaced 21 cases where formatting corrupted queries. They reduce to six root causes, all fixed here.

PostgreSQL

@name::type printed as @ name::type. The cast binds tighter than @, so the printer saw a prefix operator and added a space — and sqlc's named-param scanner then no longer matched, silently dropping the parameter from generated code. Prefix sigils (@, :, $) now glue to operands that start with an identifier. The pg fingerprint could never catch this one (@name and @ name parse to the same tree), so the printer fix is the fix. Affected corpus cases: named_param, unnest, unnest_star.

MySQL (dolphin)

The formatter reprinted from the compiler's normalized AST, which loses information the compiler doesn't need but a formatter must keep:

• Identifier case folding — `Event` became event, renaming the table on servers where table names are case-sensitive. sqlc fmt now uses dolphin.新建FormatParser(), which preserves the author's identifier case; the compiler's parser keeps lowercasing for case-insensitive catalog matching, so analysis is unchanged.
• Literal corruption — NULL printed as '', 1.0 as 0, true as 1. Literals now convert by datum kind (ast.Null, ast.Boolean, decimals keep their written scale). This also fixes type inference: a bool literal column is now bool, not int32 (selectstatic/mysql golden regenerated to match PostgreSQL's inference).
• Dropped clauses — SELECT DISTINCT lost its DISTINCT; a UNION's trailing ORDER BY vanished; GROUP_CONCAT(... ORDER BY ...) lost its ordering; optimizer hints (/*+ ... */) were deleted; and a multi-table UPDATE ... JOIN lost its ON condition and SET qualifiers — printing an update of the whole cross product. All preserved now, via converter fixes plus two print-fidelity AST fields (UpdateStmt.TableRefs, ResTarget.Relation) that analysis ignores.
• DDL type attributes — DECIMAL(10,5) printed as DECIMAL and UNSIGNED was dropped in column definitions.
• SHOW WARNINGS rewritten — dolphin converts SHOW into a synthetic SELECT for analysis, and the printer printed that internal form.

新建 safety net: a MySQL fingerprint. dolphin now implements Fingerprint (the proof fmt already requires from PostgreSQL via pg_query): the original and formatted statements are parsed with marino and their canonical restore forms compared, with identifier case preserved and redundant parentheses unwrapped. Any statement the printer cannot provably reproduce — SHOW WARNINGS, comma-join multi-table UPDATEs — is left exactly as written, which is fmt's documented contract. This is what fixes the SHOW class, and it structurally prevents this whole family of bugs from corrupting queries again.

Testing

• Regression queries for every fixed class added to internal/endtoend/testdata/fmt/{mysql,postgresql} (goldens regenerated).
• Full --tags=examples suite passes against live PostgreSQL 16 and MySQL 9.
• Re-running the corpus verification after the fixes reports zero corrupted cases: 456 configs unchanged by fmt (including the statements it now declines to format), 344 reformatted with regenerated code and parse trees proven equivalent.

One pre-existing cosmetic limitation, unchanged here: a MySQL CREATE TABLE in a query file collapses to one line, because marino exposes no column positions for the author-line-break machinery to anchor on.

🤖 Generated with Claude Code

https://claude.ai/code/session_01UZBvADuHMZEjoF551NccgP

---

Generated by Claude Code