rebase: target the official vLLM 0.31.0 release

Use upstream release images at db9527a46873, preserve custom processor conditioning, and include weekly jobs in full Buildkite release validation.

Signed-off-by: tzhouam <tzhouam@connect.ust.hk>

tests/model_executor/stage_input_processors/test_qwen3_omni_streaming_helpers.py

@@ -986,6 +986,27 @@ def test_cosyvoice3_full_payload_replace_keys_present() -> None:
     )
 
 
+def test_cosyvoice3_conditioning_replaces_snapshot_while_audio_accumulates() -> None:
+    from vllm_omni.model_executor.stage_input_processors.cosyvoice3 import _FULL_PAYLOAD_REPLACE_KEYS
+    from vllm_omni.outputs.mm_outputs import MultimodalPayload
+    from vllm_omni.outputs.output_modality import OutputModality
+
+    previous = {key: torch.ones(1, 174) for key in _FULL_PAYLOAD_REPLACE_KEYS}
+    padded = {key: torch.zeros(1, 1500) for key in _FULL_PAYLOAD_REPLACE_KEYS}
+    previous["embed.speech_token_len"] = torch.tensor([174])
+    padded["embed.speech_token_len"] = torch.tensor([174])
+    payload = MultimodalPayload.from_dict({**previous, "audio": torch.ones(2)})
+    incoming = MultimodalPayload.from_dict({**padded, "audio": torch.ones(3)})
+    assert payload is not None and incoming is not None
+
+    payload = payload.merged_with(incoming)
+    payload.consolidate_tensors(OutputModality.AUDIO)
+
+    for key in _FULL_PAYLOAD_REPLACE_KEYS:
+        assert torch.equal(payload[key], padded[key])
+    assert torch.equal(payload["audio"], torch.ones(5))
+
+
 def test_ming_flash_omni_thinker2talker_token_only_smoke() -> None:
     """Smoke: ming_flash_omni token-only carries voice metadata."""
     from vllm_omni.model_executor.stage_input_processors.ming_flash_omni import (