diff --git a/src/processing/flink_jobs/stream_processor.py b/src/processing/flink_jobs/stream_processor.py
index dd8e465..5da5e2b 100644
--- a/src/processing/flink_jobs/stream_processor.py
+++ b/src/processing/flink_jobs/stream_processor.py
@@ -75,8 +75,17 @@ class ValidateAndEnrich(ProcessFunction):
     """
 
     def process_element(
-        self, value: str, ctx: ProcessFunction.Context
-    ) -> Iterator[tuple[str, str]]:
+        self,
+        value: str,
+        ctx: ProcessFunction.Context,  # noqa: ARG002 - required by the ProcessFunction interface
+    ) -> Iterator[tuple[str | OutputTag, str]]:
+        # PyFlink routes side outputs by *yielding* ``(OutputTag, value)`` — the
+        # Java ``ctx.output(tag, value)`` API does not exist on pyflink's
+        # ProcessFunction context (``InternalProcessFunctionContext`` has no
+        # ``.output``), so emitting via ctx raised AttributeError and broke the
+        # dead-letter path. The main output yields ``(event_id, payload)``; the
+        # framework tells them apart by the first element's type (str vs
+        # OutputTag), so the tuple main output is unambiguous. (R4 follow-up)
         from datetime import UTC, datetime
 
         from src.processing.transformations.enrichment import (
@@ -91,7 +100,7 @@ def process_element(
         try:
             event = json.loads(value)
         except json.JSONDecodeError as e:
-            ctx.output(
+            yield (
                 DEAD_LETTER_TAG,
                 json.dumps(
                     {
@@ -112,7 +121,7 @@ def process_element(
                 # available so tenant resolution sees the prefixed topic.
                 event = normalize_debezium_event(event, topic=event.get("topic"))
         except ValueError as e:
-            ctx.output(
+            yield (
                 DEAD_LETTER_TAG,
                 json.dumps(
                     {
@@ -132,7 +141,7 @@ def process_element(
         # 2. Schema validation (Pydantic models)
         schema_result = validate_event(event)
         if not schema_result.is_valid:
-            ctx.output(
+            yield (
                 DEAD_LETTER_TAG,
                 json.dumps(
                     {
@@ -160,7 +169,7 @@ def process_element(
                 if i.severity == "error"
             ]
             if error_issues:
-                ctx.output(
+                yield (
                     DEAD_LETTER_TAG,
                     json.dumps(
                         {
diff --git a/tests/unit/test_stream_processor.py b/tests/unit/test_stream_processor.py
index 8b5c20b..ec32932 100644
--- a/tests/unit/test_stream_processor.py
+++ b/tests/unit/test_stream_processor.py
@@ -141,11 +141,14 @@ def execute(self, job_name):
 
 
 class _FakeProcessContext:
-    def __init__(self):
-        self.outputs = []
+    """Stand-in for pyflink's ProcessFunction.Context.
 
-    def output(self, tag, value):
-        self.outputs.append((tag, value))
+    Real pyflink has **no** ``ctx.output()`` — that is the Java side-output API
+    (``InternalProcessFunctionContext`` lacks it). Side outputs are emitted by
+    *yielding* ``(OutputTag, value)``. The fake deliberately omits ``output`` so
+    a regression back to ``ctx.output(...)`` fails loudly here, not only on the
+    live cluster.
+    """
 
 
 class _ValidationResult:
@@ -522,9 +525,11 @@ def test_process_element_routes_corrupt_json_to_dlq(stream_processor, monkeypatc
 
     emitted = list(processor.process_element("{bad json", ctx))
 
-    assert emitted == []
-    assert len(ctx.outputs) == 1
-    tag, payload = ctx.outputs[0]
+    # Invalid events are routed to the dead-letter side output by *yielding*
+    # (DEAD_LETTER_TAG, payload); the framework distinguishes side from main
+    # output by the first element's type. Nothing else is emitted.
+    assert len(emitted) == 1
+    tag, payload = emitted[0]
     assert tag is stream_processor.DEAD_LETTER_TAG
     assert json.loads(payload)["stage"] == "parse"
 
@@ -542,8 +547,10 @@ def test_process_element_routes_schema_errors_to_dlq(stream_processor, monkeypat
 
     emitted = list(processor.process_element(json.dumps({"event_type": "order.created"}), ctx))
 
-    assert emitted == []
-    assert json.loads(ctx.outputs[0][1]) == {
+    assert len(emitted) == 1
+    tag, payload = emitted[0]
+    assert tag is stream_processor.DEAD_LETTER_TAG
+    assert json.loads(payload) == {
         "event_id": "unknown",
         "error": [{"type": "missing", "field": "event_id"}],
         "stage": "schema_validation",
@@ -580,7 +587,7 @@ def _validate_event(event):
     pairs = list(processor.process_element(value, ctx))
     emitted = [json.loads(payload) for _, payload in pairs]
 
-    assert ctx.outputs == []
+    assert all(tag is not stream_processor.DEAD_LETTER_TAG for tag, _ in pairs)
     assert validated_events[0]["event_type"] == "order.created"
     assert validated_events[0]["operation"] == "insert"
     assert emitted[0]["source"] == "postgres_cdc"
@@ -608,8 +615,10 @@ def test_process_element_routes_semantic_errors_to_dlq(stream_processor, monkeyp
 
     emitted = list(processor.process_element(value, ctx))
 
-    assert emitted == []
-    assert json.loads(ctx.outputs[0][1]) == {
+    assert len(emitted) == 1
+    tag, payload = emitted[0]
+    assert tag is stream_processor.DEAD_LETTER_TAG
+    assert json.loads(payload) == {
         "event_id": "evt-1",
         "error": [
             {
@@ -644,9 +653,10 @@ def test_process_element_ignores_semantic_warnings(stream_processor, monkeypatch
         }
     )
 
-    emitted = [json.loads(payload) for _, payload in processor.process_element(value, ctx)]
+    pairs = list(processor.process_element(value, ctx))
+    emitted = [json.loads(payload) for _, payload in pairs]
 
-    assert ctx.outputs == []
+    assert all(tag is not stream_processor.DEAD_LETTER_TAG for tag, _ in pairs)
     assert emitted[0]["enriched_by"] == "payment"