### What happened?

When a streamed response is reassembled by `stream_chunk_builder`, `thinking_blocks` reasoning text is silently dropped whenever a thinking block's **signature never arrives**.

`ChunkProcessor.get_combined_thinking_content` in `litellm/litellm_core_utils/streaming_chunk_builder_utils.py` only emits a block when both accumulated text **and** a signature are present:

```python if len(current_thinking_text_parts) > 0 and current_signature: # requires signature ```

The per-block `signature` only arrives in a dedicated delta at the **end** of a thinking block. Any block whose signature never lands — a stream truncated mid-thinking by `max_tokens` or client cancellation, or a provider/gateway that emits thinking text without per-block signatures — has all of its accumulated reasoning text discarded, and `thinking_blocks` comes back as `None`.

### Minimal reproduction (offline, no network/keys)

```python from litellm import stream_chunk_builder from litellm.types.utils import Delta, ModelResponseStream, StreamingChoices

def mk(delta, fr=None): return ModelResponseStream(id="x", created=1, model="claude-x", object="chat.completion.chunk", choices=[St…