docs(score-centering): name the router that supports the recipe

examples/infra_features/score_centering/README.md

@@ -21,8 +21,6 @@ Add these arguments to an existing text-only GRPO training recipe:
 --calculate-per-token-loss
 ```
 
-Unfiltered session-server rollouts with more than 20 candidates, and filtered rollouts, need a compatible SGLang router build; see [Rollout and data contract](#rollout-and-data-contract).
-
 Keep reward mean subtraction enabled. Disabling standard-deviation normalization gives the paper's group-centered rewards. This is a separate REINFORCE-style loss: PPO clipping parameters do not apply. Existing batch size and update scheduling still control how many updates consume a rollout batch; choose them explicitly when reproducing an experiment.
 
 Choose `--score-centering-is tis` for weights clipped at `--score-centering-tis-clip` (default 2). Choose `mis` to retain ratios in `[--score-centering-mis-low, --score-centering-mis-high]` (defaults 0.5 and 5), setting other weights to zero. These weights are centered together with the score. Use these options instead of `--use-tis` or a custom TIS function.
@@ -82,8 +80,6 @@ Native SGLang generation, the legacy rollout path, and both session-server versi
 
 Evaluation requests skip this collection and may use independent sampling settings, including greedy decoding. The built-in agentic producer marks evaluation sessions when creating them; custom session clients should create them with `POST /sessions` with JSON body `{"evaluation": true}`.
 
-Unfiltered session rollouts with more than 20 candidates need an SGLang router build that accepts the requested OpenAI `top_logprobs` value and preserves `input_ids` and `return_meta_info`. Older builds cap `top_logprobs` at 20. Verify a chat request through the installed router before launching; for `--rollout-top-logprobs-num 128`, confirm that 128 candidate log probabilities reach the session server. Native `/generate` requests use `top_logprobs_num` and do not share that chat validation limit. Filtered rollouts request support probabilities through `sampling_logprobs_mode` instead of `top_logprobs`, so that cap does not apply, but the router must forward `sampling_logprobs_mode` on chat and `/generate` requests.
-
 Unused candidate slots and non-trained observation rows contain token ID `-1` and log probability `-inf`. Tool-observation masks, multi-turn merging, retries, trailing-token trimming, and truncation preserve row alignment. Session serialization retains both arrays. For score centering in support mode, training batches share the recorded support IDs when every sample has exactly the same candidate order and prefix padding. A per-row candidate count preserves observation rows and masked generated rows; any mismatch keeps the original arrays for the whole batch. The trainer reconstructs only its context-parallel rows. The source Samples and session payloads still retain both representations.
 
 Custom rollout producers must supply these fields with probabilities from the actual generation call and no repeated non-negative token ID in a row. Conversion always rejects missing candidate IDs, candidate log probabilities, or sampled-token log probabilities, including from custom producers. With `--ci-test`, Miles also validates the complete sample before training, rejecting duplicate candidates, sampling-support mismatches, and disagreeing sampled/candidate probabilities. This full-sample validation is skipped in normal training because sorting and support matching are expensive on long responses. Rescoring old rollouts with newer weights is not a substitute. With the feature disabled, requests and the session wire format remain unchanged.
@@ -93,6 +89,7 @@ Custom rollout producers must supply these fields with probabilities from the ac
 - The shared loss is wired into Megatron and FSDP. Candidate selection supports tensor parallelism, packed (`thd`) and padded (`bshd`) zigzag context parallelism, and packed all-gather context parallelism.
 - Sampling requires a fixed positive temperature and `min_p=0` on every call. Filtered sampling, including top-p filtering, automatically uses support mode and requires a positive `top_k`, for example `--rollout-top-p 0.9 --rollout-top-k 64 --rollout-top-logprobs-num 128`. Global filtered rollout settings automatically enable sampling-support replay; per-request overrides are checked when each request is built. See the [sampling-support replay guide](../../../docs/advanced/sampling-support-replay.md) for its request and server requirements.
 - Filtered sampling requires SGLang with support log probabilities (SGLang PR [#40932](https://github.com/sgl-project/sglang/pull/40932), included in the `sglang-miles` branch by [#41047](https://github.com/sgl-project/sglang/pull/41047), merge commit [`ae04cb14046896b6d453758c5769d639deedd353`](https://github.com/sgl-project/sglang/commit/ae04cb14046896b6d453758c5769d639deedd353) or a descendant containing it). External servers must set `SGLANG_RETURN_ORIGINAL_LOGPROB=0` like Miles-managed workers. OpenAI session responses must expose the SGLang fields above in `choices[0].meta_info`; a generic OpenAI-compatible server without that metadata is insufficient.
+- The SGLang router must include [sgl-router-for-miles #21](https://github.com/radixark/sgl-router-for-miles/pull/21) (`df2c790`; see `Git Commit` in `smg --version-verbose`), which raises the chat `top_logprobs` limit from 20 to 128 and forwards `sampling_logprobs_mode`.
 - Constrained/custom sampling, speculative decoding, true-on-policy mode, OPD, multi-LoRA/Tinker losses, sequence masking, custom policy-loss reducers, custom train-data converters, and logprob recomputation via prefill are rejected. Multimodal token expansion is not supported. The initial advantage estimator is GRPO.
 - Retaining `k=128` uses about 1 KiB per response position for the two arrays, before transport overhead. Eligible support-mode training batches replace the duplicate ID array with one int32 count per response position, saving `4 * (k - 1)` bytes per position in training transport and storage. Source Samples and session payloads keep the original storage cost. Larger `k` improves the unfiltered tail approximation at additional storage and compute cost.
 
@@ -125,4 +122,4 @@ MILES_LIVE_SCORE_CENTERING_SERVED_MODEL=your-served-model \
 python -m pytest --confcutdir=tests/manual tests/manual/test_score_centering_live.py
 ```
 
-Set `SGLANG_RETURN_ORIGINAL_LOGPROB=0` on the server before starting it. The probe checks native and OpenAI response metadata using the production candidate collector and validator, and checks temperature scaling at 0.7, 1.0 and 1.3. It warms the shared prompt first so that cached and uncached prefills do not confound the temperature comparison. Set `MILES_LIVE_SCORE_CENTERING_ARTIFACT_DIR` to keep the raw responses.
+Set `SGLANG_RETURN_ORIGINAL_LOGPROB=0` on the server before starting it. The probe checks unfiltered native and OpenAI response metadata using the production candidate collector and validator, and checks temperature scaling at 0.7, 1.0 and 1.3. It warms the shared prompt first so that cached and uncached prefills do not confound the temperature comparison. Set `MILES_LIVE_SCORE_CENTERING_ARTIFACT_DIR` to keep the raw responses.