diff --git a/ocr-service/test_metrics.py b/ocr-service/test_metrics.py
index 253cae14..d2bd9671 100644
--- a/ocr-service/test_metrics.py
+++ b/ocr-service/test_metrics.py
@@ -501,7 +501,17 @@ async def test_ocr_processing_seconds_histogram_observed_per_page_in_guided_stre
 
 @pytest.mark.asyncio
 async def test_ocr_processing_seconds_histogram_excludes_spell_check_time_in_guided_stream(fresh_metrics):
-    """The guided observation must time engine work only, not the spell-check pass."""
+    """The guided observation must time engine work only, not the spell-check pass.
+
+    Wall-clock bound rather than a structural `patch("main.time.monotonic")`:
+    the patched attribute is the *global* `time.monotonic`, which httpx and
+    asyncio also consume — they exhaust the deterministic sequence before the
+    request reaches the engine loop. Bound is sized against the failure mode,
+    not the noise floor: spell-check sleeps 0.05s × 2 regions = 0.1s, so a
+    timer that accidentally wrapped `correct_text` would observe >= 0.1s. The
+    0.09s ceiling catches that bug while leaving ~90ms of slack for slow CI
+    runners (engine work is instantaneous under the mock).
+    """
     mock_images = [Image.new("RGB", (100, 100))]
     regions = [
         {"pageNumber": 1, "x": 0.0, "y": 0.0, "width": 0.5, "height": 0.5, "annotationId": "a1"},
@@ -532,10 +542,7 @@ async def test_ocr_processing_seconds_histogram_excludes_spell_check_time_in_gui
     sum_seconds, _ = _histogram_count_sum(
         fresh_metrics.ocr_processing_seconds, engine="kraken"
     )
-    # Spell-check sleeps 0.05s per region × 2 regions = 0.1s; engine work is instantaneous.
-    # If timing included spell-check, sum_seconds would be >= 0.1s. Allow 30ms slack
-    # for scheduler overhead.
-    assert sum_seconds < 0.05, f"timing must exclude spell-check; got sum={sum_seconds}"
+    assert sum_seconds < 0.09, f"timing must exclude spell-check; got sum={sum_seconds}"
 
 
 @pytest.mark.asyncio