2.0.2

Fix: a stray <think> no longer discards the turn it appears in. A thinking turn is
stored so that re-rendering it extends the live KV cache instead of re-prefilling the
transcript. Two helpers held that invariant, and both assumed at most one, well-formed
think block — so a model that wrote <think> inside its own reasoning defeated both at
once: the block was never closed, the turn was stored raw and re-rendered with an
injected empty block (diverging from the cache at the turn's first token), and the
reasoning before the stray tag was dropped from the transcript entirely.

Now only the closing tag decides whether a block is open, and only a leading <think>
is treated as a delimiter — any later tag is reasoning the model wrote, not structure.
Measured on the Qwen3.8 template, a stray-tag turn re-prefilled 1146 tokens before the
fix and 22 after, with the dropped reasoning restored.