diff --git a/src/mcaxl/route_plan.py b/src/mcaxl/route_plan.py
index e11f6d3..ab284b9 100644
--- a/src/mcaxl/route_plan.py
+++ b/src/mcaxl/route_plan.py
@@ -1734,22 +1734,7 @@ def translation_chain(client: "AxlClient", number: str, css_name: str | None = N
         """
 
     # Pull every pattern in scope; we filter in Python with wildcard logic.
-    sql = f"""
-    SELECT
-      np.dnorpattern AS pattern,
-      tpu.name AS pattern_type,
-      rp.name AS partition_name,
-      np.callingpartytransformationmask AS calling_party_xform_mask,
-      np.calledpartytransformationmask AS called_party_xform_mask,
-      np.prefixdigitsout AS prefix_digits_out,
-      ddi.name AS digit_discard_instructions,
-      rf.name AS route_filter,
-      np.description AS description
-    FROM numplan np
-    LEFT OUTER JOIN routepartition rp ON np.fkroutepartition = rp.pkid
-    LEFT OUTER JOIN typepatternusage tpu ON np.tkpatternusage = tpu.enum
-    LEFT OUTER JOIN digitdiscardinstruction ddi ON np.fkdigitdiscardinstruction = ddi.pkid
-    LEFT OUTER JOIN routefilter rf ON np.fkroutefilter = rf.pkid
+    #
     # tkpatternusage candidates included in the match scan:
     #   2 = Device (DN)        — directly-dialable directory numbers
     #   3 = Translation        — translation patterns
@@ -1766,6 +1751,22 @@ def translation_chain(client: "AxlClient", number: str, css_name: str | None = N
     # was the cause of the false-positive: candidates_evaluated stayed
     # 23 vs 26 numplan rows in the reachable partition, with the 3-row
     # gap being exactly the 3 Device DNs.
+    sql = f"""
+    SELECT
+      np.dnorpattern AS pattern,
+      tpu.name AS pattern_type,
+      rp.name AS partition_name,
+      np.callingpartytransformationmask AS calling_party_xform_mask,
+      np.calledpartytransformationmask AS called_party_xform_mask,
+      np.prefixdigitsout AS prefix_digits_out,
+      ddi.name AS digit_discard_instructions,
+      rf.name AS route_filter,
+      np.description AS description
+    FROM numplan np
+    LEFT OUTER JOIN routepartition rp ON np.fkroutepartition = rp.pkid
+    LEFT OUTER JOIN typepatternusage tpu ON np.tkpatternusage = tpu.enum
+    LEFT OUTER JOIN digitdiscardinstruction ddi ON np.fkdigitdiscardinstruction = ddi.pkid
+    LEFT OUTER JOIN routefilter rf ON np.fkroutefilter = rf.pkid
     WHERE np.tkpatternusage IN (2, 3, 5, 7)
       AND np.dnorpattern IS NOT NULL
       {css_filter}
diff --git a/tests/test_cti_failsafe_reachability.py b/tests/test_cti_failsafe_reachability.py
index f51eced..27a2636 100644
--- a/tests/test_cti_failsafe_reachability.py
+++ b/tests/test_cti_failsafe_reachability.py
@@ -466,6 +466,39 @@ class TestDeviceDnInTranslationChainCandidates:
             "empirical proof."
         )
 
+    def test_no_python_comment_chars_leak_into_sql(self):
+        """Sentinel — a `#` in any captured query is almost certainly a
+        Python comment that escaped its f-string. Informix doesn't use
+        `#` for comments (it uses `--` and `/* */`); CUCM's data
+        dictionary doesn't use `#` in table or column names either.
+
+        Caught a real regression in 2026-05-09 cti-audit-prompts
+        thread: a Python explanatory comment was placed *inside* the
+        translation_chain f-string (after the JOIN clauses, before
+        WHERE). Informix returned "A syntax error has occurred" on
+        every live call. Offline tests passed because the FakeAxlClient
+        dispatched on substring matches and didn't parse the SQL.
+
+        This test wouldn't have caught the original bug if it predated
+        the fix (the FakeAxlClient still wouldn't fail) — but adding
+        it now means a future contributor can't reintroduce the same
+        class of mistake in any cti_failsafe_reachability call path.
+        """
+        client = FakeAxlClient(cti_rp_rows=[
+            _cti_row("Test-RP", "Generic", cfna="912", cfna_css="SomeCSS"),
+        ])
+        cti_failsafe_reachability(client)
+        for q in client.queries:
+            # Allow `#` in column-comment-like positions in SELECT lists?
+            # No — CUCM's data dictionary has no such columns. A `#`
+            # anywhere in any query my tool emits is a defect.
+            assert "#" not in q, (
+                f"Python `#` character leaked into SQL — likely a Python "
+                f"comment inside an f-string. Informix will reject this "
+                f"with 'A syntax error has occurred'. Offending query:\n"
+                f"{q[:200]}..."
+            )
+
     def test_cti_rp_to_cti_rp_failsafe_does_not_false_positive(self):
         """The motivating Bingham case: 911-CTI-RP CFNA → 912 (Device DN
         of 912-CTI-RP) under 911CER-CSS reaching 911CER-PT.