Knowledge-base cancellation is ignored once PDF extraction starts

Summary

prepareKnowledgeBase() checks its AbortSignal during path discovery and before staging each document, but the signal is not passed into extractPdf().

Once a PDF has been read and pdf.js parsing begins, the extraction loop processes every page without another cancellation check:

const document = await loadingTask.promise;
const pages: string[] = [];
for (let number = 1; number <= document.numPages; number++) {
  const content = await (await document.getPage(number)).getTextContent();
  pages.push(...);
}

Impact

Canceling a scan while a large or expensive PDF knowledge-base document is being parsed does not stop that work promptly. The caller remains blocked in knowledge-base preparation until extraction finishes or fails, even though cancellation is already part of this API and is honored elsewhere in the same function.

The staging directory is eventually cleaned up by the existing outer error path, but CPU/time can continue being spent after cancellation.

Expected behavior

Propagate the signal into PDF extraction and check it after loading the document and between page operations, so cancellation stops before processing further pages and still destroys the pdf.js loading task in finally.

Suggested fix

Pass signal to extractPdf(), call signal?.throwIfAborted() before each page (and after page text extraction), and add a regression that aborts on the first PDF-extraction-specific signal check. That regression should fail on current main, where only the outer preparation checks occur.

Start by tracing prepareKnowledgeBase() into extractPdf(), focusing on the existing AbortSignal checks and the pdf.js loading-task cleanup path. Add the regression described in the issue, then verify that aborting at the first PDF-extraction-specific check stops further page processing while the loading task is still destroyed in finally.