docs(golden-al): document SGLang and TRT-LLM synthetic acceptance knobs

[functionstackx]

Summary

The golden AL fairness guidelines only showed the vLLM way of injecting the committed golden acceptance length (synthetic_acceptance_length). This PR documents the equivalents for the other two engines, as requested in #2309 (review):

• golden_al_distribution/README.md (+ README_zh.md kept in parity):
  • SGLang: SGLANG_SIMULATE_ACC_LEN: '<AL>' with SGLANG_SIMULATE_ACC_METHOD: match-expected and SGLANG_SIMULATE_ACC_TOKEN_MODE: real-draft-token, set in the server environment (aggregated_environment / decode_environment in srt-slurm YAMLs, or exported in benchmark scripts) — matching existing DSv4 agentic recipes.
  • TensorRT-LLM: TLLM_SPEC_DECODE_FORCE_NUM_ACCEPTED_TOKENS, with the off-by-one called out: it counts accepted draft tokens only (excludes the bonus/verification token), so it must be set to golden AL minus 1 — e.g. AL 3.5 → 2.5. Fractional semantics (integer part always accepted, fractional part = probability of one extra draft token) verified against the linked TRT-LLM source.
• .github/codeowner-signoff-verify-prompt.md (Check 10):
  • (a) adds a TRT-LLM bullet with the AL−1 rule alongside the existing SGLang/vLLM checks;
  • (b) instructs the verifier to compare the pinned TRT-LLM value plus 1 against the golden curve, so correct configs don't false-fail.

Test plan

• Docs-only change — rendered markdown checked locally
• Env var names/values cross-checked against existing agentic recipes (benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/*.yaml) and TRT-LLM source

🤖 Generated with Claude Code

[claude]

LGTM — docs-only change, no code paths affected.

Reviewed the new SGLang and TRT-LLM synthetic-acceptance documentation in golden_al_distribution/README.md/README_zh.md and the corresponding Check 10 updates in the codeowner-signoff verify prompt. The TRT-LLM AL-1 rule is applied consistently in both places (doc example and the verifier's compare-plus-1 instruction), and the Chinese translation stays in parity with the English source.

Extended reasoning...

Overview

Docs-only PR adding SGLang and TensorRT-LLM equivalents to the vLLM synthetic-acceptance-length documentation in golden_al_distribution/README.md (+ Chinese parity file), plus a matching update to Check 10 of .github/codeowner-signoff-verify-prompt.md (an LLM-prompt template consumed by a verification workflow, not executable code).

Security risks

None. No code, config, CI logic, or permissions are touched — only markdown prose and an LLM prompt template used for downstream sign-off audits.

Level of scrutiny

Low. This is exactly the kind of small, self-contained, mechanical documentation change that does not need human review: it adds a bullet describing an env var and its off-by-one semantics, cross-checked against the linked TRT-LLM source, and keeps the two README files in parity. The bug-hunting system found no issues, and the one candidate (TRT-LLM AL-1 rule disabling forcing at golden AL=1.0) was already investigated and refuted.

Other factors

The doc claim (AL minus 1) and the verify-prompt claim (pinned value plus 1 should equal golden AL) are algebraically consistent with each other, and the fractional-acceptance semantics described match the linked TRT-LLM source referenced in the PR description. No outstanding review comments in the timeline to address.