stagehand-python@4.2.0a0.dev1541: fix(extension): read text() from direct child text nodes ( 2728) ( 3052)

In Stagehand's composed-tree XPath parser, which activates whenever a page contains a shadow root, text() was previously evaluated using element.textContent , identical to . . This caused locator evaluations to diverge from native document.evaluate() when handling nested markup, leading to incorrect matches or silent zero-result queries (such as selecting the wrong element on .first().click() ). The fix updates packages/extension/dom/locatorScripts/xpathParser.ts to distinguish between direct child text nodes ( text() ) and the element's string-value ( . ). Specifically: - text() = 'v' is now evaluated existentially across direct child text nodes. - contains(text(), 'v') and normalize-space(text()) = 'v' collapse the node-set to the string-value of the first text node, consistent with XPath specifications. - . retains its previous behavior reading the element's full subtree string-value. - If existing locators relied on text() matching nested child text, they must be updated to use . instead.

• Review existing XPath locators in automated tests or agent workflows: if any locator relied on text() matching deeply nested subtree text, update it to use . instead.