[codex] Add golden AL distributions

[functionstackx]

Summary

• build on @qiching 's work & @haic0 's work on AL collector
• add golden acceptance-length distributions for DeepSeek V4 Pro and Qwen3.5 MTP
• add golden acceptance-length distributions for Kimi K2.5 and MiniMax-M3 EAGLE3
• preserve collection metadata and link every YAML to its source Actions run
• document the SPEED-Bench collection methodology and AgentX synthetic-acceptance policy in English and Simplified Chinese

Golden-value sources

• Qwen3.5 397B: https://github.com/SemiAnalysisAI/InferenceX/actions/runs/27317114007
• Kimi K2.5 EAGLE3: https://github.com/SemiAnalysisAI/InferenceX/actions/runs/28122195822
• MiniMax-M3 EAGLE3: https://github.com/SemiAnalysisAI/InferenceX/actions/runs/28061204145
• DeepSeek V4 Pro: https://github.com/SemiAnalysisAI/InferenceX/actions/runs/27180633016

Methodology sources

• SPEED-Bench paper and dataset: https://arxiv.org/abs/2604.09557 and https://huggingface.co/datasets/nvidia/SPEED-Bench
• InferenceX synthetic-acceptance M0/M1/M2 plan: [Tracking Issue] Synthetic Acceptance for MTP Benchmarks #1651
• InferenceX collection workflow and model collectors: [NV] Add GitHub Action to collect SPEED-Bench AL matrix #1650 and [NV] Add SpeedBench AL collectors for DSR1 / GLM-5 / Qwen3.5 / Kimi EAGLE / MiniMax-M3 EAGLE #1706
• vLLM SPEED-Bench and synthetic-acceptance support: [SpecDecode][Benchmark] Add SPEED-bench support to benchmarking CLI vllm-project/vllm#36029 and [Feat] Unified Synthetic Acceptance Rate for V1 and V2 vllm-project/vllm#40662
• multi-node synthetic-acceptance bring-up: [NV] Multi-node synthetic AL injection for MTP (dsv4 mtp2 bring-up) #1789

Validation

• confirmed every committed AL value and metadata block against its run's speedbench-reference-al artifact
• parsed all four golden files with PyYAML
• verified every relative README link resolves and both language versions contain matching link targets and section counts
• verified balanced Markdown code fences
• ran git diff --check

[claude]

Claude finished @functionstackx's task in 52s —— View job

---

Review: PR #1926 — Add golden AL distributions

• Read changed files (4 YAML + 2 README)
• Check YAML validity & metadata consistency
• Verify README links and cross-language parity
• Post review

LGTM — no blocking issues found.

This is a docs/data-only change (4 golden AL YAMLs + EN/ZH READMEs), no executable code paths. Spot checks all passed:

• Every relative link resolves: .github/workflows/speedbench-al.yml and all four referenced collectors under benchmarks/single_node/speedbench/ (dsv4_fp4_b300_vllm.sh, qwen3.5_fp4_b300_vllm.sh, kimik2.5_fp4_b300_vllm.sh, minimaxm3_fp4_b300_vllm.sh) exist on this branch.
• The README example value synthetic_acceptance_length: 3.24 is internally consistent — it matches kimik2.5_eagle3.yaml → thinking_on → draft length 4.
• All four YAMLs are well-formed (integer draft-length keys 1–8, float AL values, per-model thinking_on/thinking_off blocks) and each carries its source Actions-run URL on line 1 plus full sampling metadata.
• EN and ZH READMEs are structurally parallel (matching section count, table rows, and link targets).
• The small non-monotonic dips at higher draft lengths (e.g. dsv4 thinking_on 6→7: 2.55→2.54; qwen3.5 thinking_off 7→8: 4.73→4.71) are expected for real measured AL and not data errors.
  · codex/add-golden-al-distributions

[functionstackx]

hi @chunfangamd @haic0 @xinli-sw @qiching PTAL before i merge this PR tmw at 4pm ET