[特征] 在独立的 LLM 路径中无法控制推理/思考(截断的响应变成空的提取)

(Chinese summary) The standalone path has no thinking/reasoning switch at all: when L1 runs with a 4096 budget, the reasoning model consumes the output budget, the visible content is cut off in the middle of the JSON (finishReason=length), and the extraction yields nothing. On one day, 8 batches were produced in this way for L1 (combined with a provider quota outage, they caused the data loss described in #1395). And the fetch layer probe proved that the `thinking` in the outgoing body was undefined when the field was passed via `providerOptions.OpenAI.thinking` on the compatible path — it never reached the provider, so the value has to be injected where the request body is built. We suggest adding a per-provider (and per-layer, see #1396) `disableThinking` honored by the standalone runner, implemented at the request-body/fetch layer instead of relying on providerOptions: