feat(ocr): count words and illegible words at the OCR call sites

Walks block["words"] before apply_confidence_markers strips the list, then
increments ocr_words_total by len(words) and ocr_illegible_words_total by
the count below threshold. Same pattern in both /ocr and /ocr/stream so the
ratio illegible/words is a faithful quality signal across endpoints.

Refs #652 (AC4)

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

ocr-service/main.py

@@ -124,8 +124,13 @@ async def run_ocr(request: OcrRequest):
 
     threshold = get_threshold(script_type)
     for block in blocks:
-        if block.get("words"):
+        words = block.get("words") or []
-            block["text"] = apply_confidence_markers(block["words"], threshold)
+        if words:
+            metrics.ocr_words_total.inc(len(words))
+            metrics.ocr_illegible_words_total.inc(
+                sum(1 for w in words if w["confidence"] < threshold)
+            )
+            block["text"] = apply_confidence_markers(words, threshold)
         block.pop("words", None)
         if script_type in _SPELL_CHECK_SCRIPT_TYPES:
             block["text"] = correct_text(block["text"])
@@ -258,8 +263,13 @@ async def run_ocr_stream(request: OcrRequest):
                 )
 
                 for block in blocks:
-                    if block.get("words"):
+                    words = block.get("words") or []
-                        block["text"] = apply_confidence_markers(block["words"], threshold)
+                    if words:
+                        metrics.ocr_words_total.inc(len(words))
+                        metrics.ocr_illegible_words_total.inc(
+                            sum(1 for w in words if w["confidence"] < threshold)
+                        )
+                        block["text"] = apply_confidence_markers(words, threshold)
                     block.pop("words", None)
                     if script_type in _SPELL_CHECK_SCRIPT_TYPES:
                         block["text"] = correct_text(block["text"])

ocr-service/test_metrics.py

@@ -230,3 +230,40 @@ async def test_ocr_skipped_pages_total_incremented_when_engine_raises_for_a_page
     assert fresh_metrics.ocr_skipped_pages_total._value.get() == 1.0
     # The second page still succeeds.
     assert fresh_metrics.ocr_pages_total.labels(engine="surya")._value.get() == 1.0
+
+
+@pytest.mark.asyncio
+async def test_ocr_words_and_illegible_words_total_sum_across_blocks(fresh_metrics):
+    """Counters reflect totals summed over every block in the request.
+
+    Threshold defaults to THRESHOLD_DEFAULT (0.3) for non-Kurrent scripts. Two
+    blocks: 3 words above + 2 words below threshold across blocks.
+    """
+    mock_images = [Image.new("RGB", (100, 100))]
+    mock_blocks = [
+        {"pageNumber": 1, "x": 0.0, "y": 0.0, "width": 1.0, "height": 1.0,
+         "polygon": None, "text": "ignored",
+         "words": [{"text": "Lieber", "confidence": 0.9},
+                   {"text": "Freund", "confidence": 0.1}]},
+        {"pageNumber": 1, "x": 0.0, "y": 0.0, "width": 1.0, "height": 1.0,
+         "polygon": None, "text": "ignored",
+         "words": [{"text": "Gruss", "confidence": 0.8},
+                   {"text": "verschmiert", "confidence": 0.05},
+                   {"text": "Karl", "confidence": 0.95}]},
+    ]
+
+    with patch("main.kraken_engine.load_models"), \
+         patch("main.load_spell_checker"), \
+         patch("main._download_and_convert_pdf", new_callable=AsyncMock, return_value=mock_images), \
+         patch("main.preprocess_page", side_effect=lambda img: img), \
+         patch("main.surya_engine.extract_blocks", return_value=mock_blocks):
+        async with AsyncClient(transport=ASGITransport(app=app), base_url="http://test") as client:
+            import main as main_module
+            main_module._models_ready = True
+            try:
+                await _drive_ocr(client, script_type="TYPEWRITER")
+            finally:
+                main_module._models_ready = False
+
+    assert fresh_metrics.ocr_words_total._value.get() == 5.0
+    assert fresh_metrics.ocr_illegible_words_total._value.get() == 2.0