fix(renderers): bridge merge must not mutate prior multi_modal_data

renderers/kimi_k25.py

@@ -1183,18 +1183,20 @@ def emit_image(
             emit_text("<think></think>", -1)
 
         # Merge prev mm_data (earlier-turn images) with the new turn's items.
+        # Copy the per-modality lists (not just the outer dict) so appending
+        # below never mutates the caller's previous_multi_modal_data.
         merged_hashes: dict[str, list[str]] = (
-            dict(previous_multi_modal_data.mm_hashes)
+            {k: list(v) for k, v in previous_multi_modal_data.mm_hashes.items()}
             if previous_multi_modal_data
             else {}
         )
         merged_placeholders: dict[str, list[PlaceholderRange]] = (
-            dict(previous_multi_modal_data.mm_placeholders)
+            {k: list(v) for k, v in previous_multi_modal_data.mm_placeholders.items()}
             if previous_multi_modal_data
             else {}
         )
         merged_items: dict[str, list[dict[str, Any]]] = (
-            dict(previous_multi_modal_data.mm_items)
+            {k: list(v) for k, v in previous_multi_modal_data.mm_items.items()}
             if previous_multi_modal_data
             else {}
         )

renderers/qwen35.py

@@ -834,18 +834,20 @@ def flush_buf() -> None:
             emit_text("\n\n", -1)
 
         # Merge prev mm_data (images from earlier turns) with the new turn's.
+        # Copy the per-modality lists (not just the outer dict) so appending
+        # below never mutates the caller's previous_multi_modal_data.
         merged_hashes: dict[str, list[str]] = (
-            dict(previous_multi_modal_data.mm_hashes)
+            {k: list(v) for k, v in previous_multi_modal_data.mm_hashes.items()}
             if previous_multi_modal_data
             else {}
         )
         merged_placeholders: dict[str, list[PlaceholderRange]] = (
-            dict(previous_multi_modal_data.mm_placeholders)
+            {k: list(v) for k, v in previous_multi_modal_data.mm_placeholders.items()}
             if previous_multi_modal_data
             else {}
         )
         merged_items: dict[str, list[dict[str, Any]]] = (
-            dict(previous_multi_modal_data.mm_items)
+            {k: list(v) for k, v in previous_multi_modal_data.mm_items.items()}
             if previous_multi_modal_data
             else {}
         )

renderers/qwen3_vl.py

@@ -828,19 +828,21 @@ def render_media_content(content: Any) -> None:
         em.text("assistant\n", is_sampled=False, is_content=False)
         em.finalize()
 
-        # Merge prev mm_data with the new turn's items.
+        # Merge prev mm_data with the new turn's items. Copy the per-modality
+        # lists (not just the outer dict) so appending below never mutates the
+        # caller's previous_multi_modal_data.
         merged_hashes = (
-            dict(previous_multi_modal_data.mm_hashes)
+            {k: list(v) for k, v in previous_multi_modal_data.mm_hashes.items()}
             if previous_multi_modal_data
             else {}
         )
         merged_placeholders = (
-            dict(previous_multi_modal_data.mm_placeholders)
+            {k: list(v) for k, v in previous_multi_modal_data.mm_placeholders.items()}
             if previous_multi_modal_data
             else {}
         )
         merged_items = (
-            dict(previous_multi_modal_data.mm_items)
+            {k: list(v) for k, v in previous_multi_modal_data.mm_items.items()}
             if previous_multi_modal_data
             else {}
         )

tests/test_multimodal.py

@@ -583,6 +583,12 @@ def test_multimodal_bridge_extends_and_carries_mm_data(
     ]
 
     initial_rendered = renderer.render(initial, add_generation_prompt=True)
+    prior_mm = initial_rendered.multi_modal_data
+    prior_counts = (
+        len(prior_mm.mm_placeholders.get(modality, [])),
+        len(prior_mm.mm_items.get(modality, [])),
+        len(prior_mm.mm_hashes.get(modality, [])),
+    )
     # ``previous_completion_ids`` mirrors what a sampler would emit
     # starting AFTER the prompt's assistant role opener — i.e. the
     # response text followed by ``<|im_end|>``.
@@ -632,6 +638,19 @@ def test_multimodal_bridge_extends_and_carries_mm_data(
     hashes = bridged_mm.mm_hashes.get(modality, [])
     assert len(items) == 2 and len(hashes) == 2
 
+    # (2b) The prior turn's sidecar is unchanged — the bridge copies the
+    # per-modality lists, so the carried-forward item doesn't grow the
+    # caller's previous_multi_modal_data in place.
+    assert (
+        (
+            len(prior_mm.mm_placeholders.get(modality, [])),
+            len(prior_mm.mm_items.get(modality, [])),
+            len(prior_mm.mm_hashes.get(modality, [])),
+        )
+        == prior_counts
+        == (1, 1, 1)
+    ), f"{mm_model_name} / {modality}: bridge mutated previous_multi_modal_data"
+
     # (3) Extension contains the new turn's pad run, and its
     # placeholder offset lands inside the extension region.
     pad_id = tokenizer.convert_tokens_to_ids(kit["placeholder_token"])