Fix ValueError when formatting procedures using label syntax (bar:LOOP)

[VCourdy]

Bug

SqlFormatter::format() crashes with ValueError: str_repeat(): Argument #2 ($times) must be greater than or equal to 0 on MySQL stored procedures using label syntax with no space after the colon:

use Doctrine\SqlFormatter\NullHighlighter;
use Doctrine\SqlFormatter\SqlFormatter;

$stmt = 'CREATE PROCEDURE `foo`()
BEGIN
  bar:LOOP
    -- Nothing
  END LOOP bar;
END';

(new SqlFormatter(new NullHighlighter()))->format($stmt);
// ValueError in src/SqlFormatter.php:94

bar: LOOP (with a space) does not crash; the trigger is specifically : glued to LOOP.

Root cause

The tokenizer lexes : as a bind-variable prefix, so bar:LOOP tokenizes as WORD('bar') + VARIABLE(':LOOP'). The LOOP keyword is swallowed and the indent-opener branch in SqlFormatter::format() never fires, while END LOOP and the final END each still decrement the indentation level. It reaches -1 and str_repeat($tab, $indentLevel) throws.

Fix

A colon immediately preceded by a word character is never a bind-variable prefix in any supported dialect — it is a label separator (MySQL/Oracle label:LOOP). This PR guards the : alternative of the VARIABLE pattern with a negative lookbehind:

-Token::TOKEN_TYPE_VARIABLE => '[@:](?:...)',
+Token::TOKEN_TYPE_VARIABLE => '(?:@|(?<!\w):)(?:...)',

@ keeps its old unconditional behavior; only : gets the guard.

Impact on existing behavior

• Zero changes to existing fixtures: all pre-existing sections of tests/sql.sql render byte-identically across all five expected-output files (only the new regression section is appended).
• Bind variables are unaffected: :param, @var, WHERE x = :id still tokenize as TOKEN_TYPE_VARIABLE (covered by a new tokenizer test). PostgreSQL casts (1::text, '{}'::json) are unaffected — in :: the second : is preceded by :, not a word character.
• One intentional re-tokenization: a : glued to a preceding identifier or number (e.g. the :5 in a PostgreSQL array slice arr[1:5]) was VARIABLE and becomes BOUNDARY + WORD. No fixture covers it and rendered output is equivalent; such a colon is never a placeholder in any supported dialect.

Tests

• New tokenizer cases: bar:loop → WORD + BOUNDARY + WORD, and :param → VARIABLE.
• New format regression fixture (the procedure above), expected outputs regenerated with bin/regenerate-expected-output.
• phpunit, phpstan analyse and phpcs all pass.

Relates to #118 (formatter must never fail).

🤖 Generated with Claude Code