feat(ocr): observe ocr_processing_seconds around engine.to_thread calls

Wraps every asyncio.to_thread(engine.extract_*) call with time.monotonic()
deltas in /ocr (per document) and in both /ocr/stream generators (per page).
Streaming buckets are the useful operational signal; the non-streaming
observation is a bonus.

Refs #652 (AC5)

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

ocr-service/main.py

@@ -10,6 +10,7 @@ import re
 import shutil
 import subprocess
 import tempfile
+import time
 import zipfile
 from contextlib import asynccontextmanager
 from datetime import datetime, timezone
@@ -108,6 +109,7 @@ async def run_ocr(request: OcrRequest):
     script_type = request.scriptType.upper()
     engine_name = "kraken" if script_type == "HANDWRITING_KURRENT" else "surya"
 
+    extract_started = time.monotonic()
     if script_type == "HANDWRITING_KURRENT":
         if not kraken_engine.is_available():
             raise HTTPException(
@@ -119,6 +121,9 @@ async def run_ocr(request: OcrRequest):
     else:
         # TYPEWRITER, HANDWRITING_LATIN, UNKNOWN — all use Surya
         blocks = await asyncio.to_thread(surya_engine.extract_blocks, images, request.language)
+    metrics.ocr_processing_seconds.labels(engine=engine_name).observe(
+        time.monotonic() - extract_started
+    )
 
     metrics.ocr_jobs_total.labels(engine=engine_name, script_type=script_type).inc()
 
@@ -194,6 +199,7 @@ async def run_ocr_stream(request: OcrRequest):
                     image = await asyncio.to_thread(preprocess_page, image)
                     blocks = []
                     sender_path = request.senderModelPath if use_kraken else None
+                    page_started = time.monotonic()
                     for region in page_regions:
                         text = await asyncio.to_thread(
                             engine.extract_region_text, image,
@@ -213,6 +219,9 @@ async def run_ocr_stream(request: OcrRequest):
                             "annotationId": region.annotationId,
                         })
 
+                    metrics.ocr_processing_seconds.labels(engine=engine_name).observe(
+                        time.monotonic() - page_started
+                    )
                     total_blocks += len(blocks)
                     metrics.ocr_pages_total.labels(engine=engine_name).inc()
                     yield json.dumps({
@@ -258,9 +267,13 @@ async def run_ocr_stream(request: OcrRequest):
                 yield json.dumps({"type": "preprocessing", "pageNumber": page_idx}) + "\n"
                 image = await asyncio.to_thread(preprocess_page, image)
                 sender_path = request.senderModelPath if use_kraken else None
+                page_started = time.monotonic()
                 blocks = await asyncio.to_thread(
                     engine.extract_page_blocks, image, page_idx, request.language, sender_path
                 )
+                metrics.ocr_processing_seconds.labels(engine=engine_name).observe(
+                    time.monotonic() - page_started
+                )
 
                 for block in blocks:
                     words = block.get("words") or []