Possible fix for duplicate messages:

• Some providers send full content on text_end, and we were appending that on top of deltas, doubling the reply.
• Update the block streaming buffer logic to treat text_start/text_end content as authoritative and replace buffers when it matches or extends the current buffer, instead of always appending.
• Add a regression test that simulates a full-content text_end and asserts only one block reply is emitted.
• now handles both single message and multi-chunk

(coded by codex, tested locally)