feat: brain PDF/image text extraction — pymupdf + tesseract OCR + vision API

- PDF: extracts selectable text via pymupdf, falls back to Tesseract OCR for scanned docs
- PDF: renders first page as screenshot thumbnail
- Images: Tesseract OCR for text extraction, OpenAI vision API fallback for photos
- Plain text files: direct decode
- All extracted text stored in extracted_text field for search/embedding
- Tested: PDF upload → text extracted → AI classified → searchable

New deps: pymupdf, pytesseract, Pillow
System dep: tesseract-ocr added to both Dockerfiles

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

services/brain/app/worker/tasks.py

@@ -99,6 +99,66 @@ async def _process_item(item_id: str):
                         )
                         db.add(asset)
 
+            # ── Step 1b: Process uploaded files (PDF, image, document) ──
+            if item.type in ("pdf", "image", "document", "file"):
+                from app.services.extract import extract_text_from_file, describe_image_with_vision
+                from app.services.storage import storage as file_storage
+
+                # Find the original upload asset
+                upload_asset = None
+                for a in item.assets:
+                    if a.asset_type == "original_upload":
+                        upload_asset = a
+                        break
+
+                if upload_asset and file_storage.exists(upload_asset.storage_path):
+                    log.info(f"Extracting text from {item.type}: {upload_asset.filename}")
+                    file_bytes = file_storage.read(upload_asset.storage_path)
+                    result = extract_text_from_file(
+                        file_bytes,
+                        upload_asset.content_type or "application/octet-stream",
+                        upload_asset.filename,
+                    )
+
+                    if result["text"]:
+                        extracted_text = result["text"]
+                        log.info(f"Extracted {len(extracted_text)} chars via {result['method']}")
+
+                    # Save PDF screenshot as an asset
+                    if result.get("screenshot_png"):
+                        from app.services.storage import storage
+                        path = storage.save(
+                            item_id=item.id,
+                            asset_type="screenshot",
+                            filename="screenshot.png",
+                            data=result["screenshot_png"],
+                        )
+                        asset = ItemAsset(
+                            id=str(uuid.uuid4()),
+                            item_id=item.id,
+                            asset_type="screenshot",
+                            filename="screenshot.png",
+                            content_type="image/png",
+                            storage_path=path,
+                        )
+                        db.add(asset)
+
+                    # For images with little OCR text, try vision API for description
+                    if item.type == "image" and len(extracted_text) < 50:
+                        log.info("Image has little OCR text, trying vision API...")
+                        vision_text = await describe_image_with_vision(
+                            file_bytes,
+                            upload_asset.content_type or "image/png",
+                        )
+                        if vision_text:
+                            extracted_text = vision_text
+                            log.info(f"Vision API returned {len(vision_text)} chars")
+
+                    item.metadata_json = item.metadata_json or {}
+                    item.metadata_json["extraction_method"] = result["method"]
+                    if result.get("page_count"):
+                        item.metadata_json["page_count"] = result["page_count"]
+
             # ── Step 2: AI classification ──
             log.info(f"Classifying item {item.id}")
             classification = await classify_item(