fix: use Google Vision OCR instead of AI-vision estimates for Windows Type Agent clicks

Claude vision's self-reported text coordinates had a small but consistent bias,
causing the patient-row double-click to land above the name and the Exam-column
divider count to occasionally miscount and hard-fail. Swap patient-row, Exam,
and Save button targeting to exact OCR bounding boxes (Google Vision via the
existing PaymentOCRService) instead of AI-estimated ratios.

apps/PaymentOCRService/main.py

@@ -5,11 +5,13 @@ from typing import List, Optional
 import io
 import os
 import asyncio
+import tempfile
 import uvicorn
 from dotenv import load_dotenv
-load_dotenv() 
+load_dotenv()
 
 from complete_pipeline_adapter import process_images_to_rows,rows_to_csv_bytes
+from complete_pipeline import extract_words_and_text
 from pdf_extractor import extract_ra_pdf
 
 app = FastAPI(
@@ -98,6 +100,41 @@ async def extract_json(files: List[UploadFile] = File(...)):
             async with lock:
                 active_jobs -= 1
 
+@app.post("/extract/words")
+async def extract_words(file: UploadFile = File(...)):
+    """Raw OCR only — no deskew/line-grouping/payment-domain logic, just the
+    flat word list with pixel bounding boxes. Used by the Windows Type Agent
+    to locate exact click targets (e.g. a patient name) instead of relying
+    on an AI-vision estimate of where text is."""
+    _validate_files([file])
+
+    async with lock:
+        global waiting_jobs
+        waiting_jobs += 1
+
+    async with semaphore:
+        async with lock:
+            waiting_jobs -= 1
+            global active_jobs
+            active_jobs += 1
+
+        tmp_path = None
+        try:
+            suffix = os.path.splitext(file.filename or "")[1] or ".png"
+            blob = await file.read()
+            with tempfile.NamedTemporaryFile(delete=False, suffix=suffix) as tmp:
+                tmp.write(blob)
+                tmp_path = tmp.name
+            words, _ = extract_words_and_text(tmp_path)
+            return JSONResponse(content={"words": words})
+        except Exception as e:
+            raise HTTPException(status_code=500, detail=f"OCR error: {e}")
+        finally:
+            if tmp_path and os.path.exists(tmp_path):
+                os.remove(tmp_path)
+            async with lock:
+                active_jobs -= 1
+
 @app.post("/extract/csvtext", response_class=PlainTextResponse)
 async def extract_csvtext(files: List[UploadFile] = File(...)):
     _validate_files(files)