refactor: consolidate AI analyzers into shared lib/ai module and add journal entry reversal columns

Extract shared vision/document analysis logic into lib/ai (vision-client, document-analyzer,
image preprocessing, validation helpers), simplifying invoice-analyzer, receipt-analyzer, and
document classifier. Add migration 046 for journal entry reversal/correction link columns
(reversed_by_id, reverses_id, correction_of_id) required by storno service. Update extension
components and shared UI.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

lib/ai/preprocess-image.ts

@@ -0,0 +1,43 @@
+/**
+ * Image Preprocessing for OCR
+ *
+ * SERVER-ONLY: Uses sharp for image processing.
+ *
+ * Preprocesses receipt/invoice images before sending to Claude Vision.
+ * Targets faded thermal prints with low contrast — the primary source
+ * of garbled OCR output.
+ */
+
+import 'server-only'
+import sharp from 'sharp'
+
+/**
+ * Preprocess an image for better OCR accuracy.
+ * Converts to grayscale, normalizes contrast, and sharpens.
+ * Returns base64-encoded JPEG.
+ *
+ * For PDFs: returns input unchanged (sharp doesn't handle PDFs).
+ */
+export async function preprocessImage(
+  base64: string,
+  mimeType: string
+): Promise<{ base64: string; mimeType: string }> {
+  // PDFs are not image files — return as-is
+  if (mimeType === 'application/pdf') {
+    return { base64, mimeType }
+  }
+
+  const buffer = Buffer.from(base64, 'base64')
+
+  const processed = await sharp(buffer)
+    .grayscale()
+    .normalize()
+    .sharpen({ sigma: 1.5 })
+    .jpeg({ quality: 90 })
+    .toBuffer()
+
+  return {
+    base64: processed.toString('base64'),
+    mimeType: 'image/jpeg',
+  }
+}