[musatoktas]

Fixes #187

The regexes in #187 are slow on a line where a rule finds its start many times but cannot complete a match. The engine tries every start again, and each try reads to the end of the line, so the time grows with the square of the line length.

What changes

• Each rule in errorMessagePrettifier.ts gets a last alternative after its start: (?<rest>.*). When a rule cannot match from a start, no later start on the same line can match either, because the content parts are .*? or .* (a match from a later start is also a match from the earlier one). So the rest group takes the rest of the line in one step. For the ['“]([^>]*)['”] (type|...) rule the rest is [^>]*, and for the last rule it is (?:(?!["]).)*: the same characters as their content.
• createErrorMessagePrettifier keeps a match with a rest group as it is.
• addMissingParentheses.ts:54: the content cannot go past a ( that starts another (name:. The match then ends at the same ), so the result is the same. In addMissingParentheses the old regex was not slow, because the function adds a ) for each unclosed (. This change only makes the regex itself linear.

Apart from that, the regexes and their captures do not change. I did not use negated character classes: TypeScript writes quoted property names with single quotes ({ 'first-name': string; }), and a class like [^']* changes the output for those (errorWithDashInObjectKeys, errorWithSpecialCharsInObjectKeys and the is missing the following properties case).

Checks

• recheck 4: before, all 13 regexes from the issue table are vulnerable (polynomial). After, all 13 are safe.

• The attack strings from the issue, through the prettifier, Node 20 on GitHub Actions:

  attack string	main (ms)	this branch (ms)
  string literal	3397	19
  declare module	7798	6
  missing properties	3465	5
  types and	3789	17
  type annotation	3817	4
  overload	3802	14
  module or file	2082	2
  type	6682	4
  quoted before a keyword	2357	2
  keyword	1409	1
  return or operator	3776	3
  quoted	4770	33

• Same output: I compared main and this branch on 8,506 messages, with the test stub and with hoverCodeBlock. The messages are the 22 mocks and the 2,121 messages of TypeScript 5.9.3's diagnosticMessages.json, each filled 4 times with different types. There are 0 differences.

• Tests: one new test per attack string, which must finish in under 500 ms. Without the source change, all 12 fail (1.4 s to 7.2 s each). With it, the 39 formatter tests and the 10 vscode-formatter tests pass, and format:check and the build pass.

This covers the 13 regexes in the issue table.