apr serve stop handling: stop on every EOS and stop string, on every route and backend #5313

Scope: Stop-token handling in apr serve: /v1/completions running past stop on CPU GGUF, GPU paths stopping on a single EOS or none, native routes leaking the im_end control token into text, and one shared sampling/stop function instead of several.

Why it matters: Clients see run-on text and control tokens; the same model stops correctly in llama-server and ollama.

Done when: On the CRUX prompt set, every apr serve route on CPU and GPU stops where llama-server stops and returns no control tokens, through one shared stop function.

Epic: #3999. Folded under triage rule 7b (2026-10-14): each row below was its own issue, closed as not planned and kept here as a sub-ticket. Reopen an issue when its work starts or its PR opens.

Sub-tickets

• apr serve GGUF CPU: POST /v1/completions runs past stop on 13/13 CRUX prompts (llama-server stops) #4339 apr serve GGUF CPU: POST /v1/completions runs past stop on 13/13 CRUX prompts (llama-server stops) (milestone 0.73.0); branch fix/4339-gguf-completions-stop
  • accept: the outcome in the title, with a receipt
• apr serve native routes leak <|im_end|> into text: /generate, /batch/generate, /stream/generate, /realize/generate, /realize/batch #4344 apr serve native routes leak <|im_end|> into text: /generate, /batch/generate, /stream/generate, /realize/generate, /realize/batch (milestone 0.71.0); branch docs/findings-3d-4344-verify
  • accept: the outcome in the title, with a receipt
• apr serve: GPU/CUDA /v1/completions paths stop on a single EOS or none (follow-up to #4339) #4345 apr serve: GPU/CUDA /v1/completions paths stop on a single EOS or none (follow-up to apr serve GGUF CPU: POST /v1/completions runs past stop on 13/13 CRUX prompts (llama-server stops) #4339) (milestone 0.73.0); branch w71-e3a/4345-stream-generate-stop
  • accept: the outcome in the title, with a receipt
• Delete dead generate paths + one shared sampling/stop-token fn (verb→engine map follow-up) #4266 Delete dead generate paths + one shared sampling/stop-token fn (verb→engine map follow-up) (milestone 0.71.0); branch fix/4266-p2-shared-sampler
  • accept: the outcome in the title, with a receipt