[Bug]: getByText/getByLabel with a global or sticky regex (/x/g, /x/y) match only some of the elements #42813

Version

1.64.0-next (main @ 07f1a61); long-standing

Steps to reproduce

import { test, expect } from '@playwright/test';

test('regex with g flag', async ({ page }) => {
  await page.setContent(`
    <div>foo</div><div>foo</div><div>foo</div>
    <label>foo<input></label><label>foo<input></label><label>foo<input></label>
  `);
  console.log(await page.getByText(/foo/).count());   // 6
  console.log(await page.getByText(/foo/g).count());  // 4
  console.log(await page.getByLabel(/foo/g).count()); // 2
  console.log(await page.getByText(/foo/y).count());  // 4
});

Expected behavior

The g and y flags have no meaning for a locator match, so /foo/g should match the same elements as /foo/. getByRole('button', { name: /foo/g }) and filter({ hasText: /foo/g }) already behave that way.

Actual behavior

Every other matching element is skipped, depending on traversal order:

6
4
2
4

Additional context

createTextMatcher in packages/injected/src/injectedScript.ts builds one RegExp and reuses re.test(elementText.full) for every element. RegExp.prototype.test is stateful when the g or y flag is set: lastIndex advances after a hit and only resets after a miss, so the next element is tested from a non-zero offset. escapeRegexForSelector passes user flags through verbatim and the selector parser accepts [dgimsuvy], so the flag reaches the engine.

Role-name matching uses String.prototype.match and has-text builds a fresh RegExp per element, which is why those are unaffected and the inconsistency is confusing. Copying a /pattern/g regex from String.replace code into getByText is an easy habit; the result is a silently wrong count or a hidden strict-mode violation.

I intend to work on this and will send a PR.

Environment

- Operating System: macOS (Darwin 25.6.0)
- Node.js: 24.8.0
- Browser: Chromium (bundled r1246)
- Playwright: main @ 07f1a6154