apr serve native routes leak <|im_end|> into text: /generate, /batch/generate, /stream/generate, /realize/generate, /realize/batch #4344

Gap from #4340 (the /v1/chat/completions stream fix, branch x/4340-stream-im-end).

Observed. apr serve --no-gpu on a Qwen GGUF, multi-turn prompt ctl-recall-7, second turn. Evidence is in /mnt/nvme-raid0/tmp/tmp.ZtvuBEgtRL/cc324af070c2/off/serve/apr/:

recall-cats shows the same leak.

Cause. The model writes the end-of-turn marker out as ordinary text tokens (<|im, _end|>), so the EOS stop token never fires. A whole special token decodes to "", which is why only the spelled-out form leaks. None of these routes truncate text at a marker.

Fix scope. These are RAW completion routes, so only special-token markers are cut (<|im_end|>, <|endoftext|>, <|end|>, </s>, <|im_start|>). The chat turn markers (\nUser:) are not cut, because a raw completion may legitimately contain them. text echoes the prompt, so the cut is searched only AFTER the prompt's decode; a chat-templated prompt carries markers of its own. The pregenerated stream stops at the token that contains the marker start and truncates that token's text.

Falsifier. One test per handler (the five routes map to three handlers).