Refresh skill validation evidence

[BBuf]

Summary

• Refresh live upstream anchors and open/merged PR watch rows for SGLang, vLLM, TensorRT-LLM, and TokenSpeed as of 2026-06-27.
• Register model-pr-diff-dossier in the top-level install/reload flow and add a README quality test so it does not drift out of the public skill list again.
• Tighten TensorRT-LLM and TokenSpeed model-history cards with explicit file-trace commands, PR-files API provenance, diff details, and risk/validation fields.
• Add the verified SGLang LTX2 residual-gate fusion pattern from merged PR [KDA-Pilot] Add diffusion residual-gate CUDA fast path for LTX2 sgl-project/sglang#29361 to the torch-profiler catalog/helper.

Live source checks

• SGLang: sgl-project/sglang@9214b9338fcb1df5b1a9ab5835aa631539f55556
• vLLM: vllm-project/vllm@091d13976c1c246714bb2112dd2e208561dda6a3
• TensorRT-LLM: NVIDIA/TensorRT-LLM@aaffa2f9fef3025e0f698d978385a73460344e0b
• TokenSpeed: lightseekorg/tokenspeed@d0a7faddb5ec0d4c6d037c4c3e6a781d2c5164a8

I also checked current open/recent PRs for Qwen3.5, DeepSeek-V4, Kimi K2.5/KimiLinear, GLM-5/5.2, MoE, GDN/KDA/MLA, and serving benchmark relevance. This PR updates top-level watch rows and source anchors, but intentionally does not rewrite the entire per-model history corpus; a dry-run rebuild showed a very large full-card rewrite, so I kept this scoped to validation evidence and stale/incorrect public guidance.

Validation

Local:

• python3 -m pytest -q -> 111 passed in 4.42s
• python3 -m black --check tests/test_model_pr_dossier_quality.py
• python3 -m py_compile skills/llm-torch-profiler-analysis/scripts/triage_kernel_helpers.py skills/llm-serving-auto-benchmark/scripts/validate_cookbook_configs.py skills/model-compute-simulation/scripts/model_compute_simulator.py skills/model-architecture-diagram/scripts/model_architecture_diagram.py
• python3 skills/llm-serving-auto-benchmark/scripts/validate_cookbook_configs.py skills/llm-serving-auto-benchmark/configs/cookbook-llm
• Five-model compute simulator smoke: qwen3-235b-a22b, deepseek-v4-flash, kimi-k2.5, minimax-m2, glm-5
• Five-model architecture diagram smoke: Qwen3-235B-A22B, Kimi K2.5, DeepSeek V4, GLM-5, MiniMax M2
• SGLang humanize corpus smoke: scanned 32639 threads, matched 296 threads across 132 PRs
• Model-history query smoke for TokenSpeed Qwen3.5 paths

ion-b200:

• Remote temp tree under /tmp/ai_skills_validate
• pytest -> 111 passed, 132 subtests passed in 9.16s
• Cookbook validation + compute simulator + architecture diagram smoke -> passed
• py_compile for relevant scripts -> passed
• Nsight Compute smoke with ncu --set basic -> passed
• Temporary remote validation files cleaned up after the run

[BBuf]

PR 72 updated with a full skill-refresh pass.

Highlights:

• Refreshed SGLang/vLLM model PR histories to live heads: SGLang e0c0c0a45cb1bda90392bfa2bba4184f5b0638a0, vLLM 091d13976c1c246714bb2112dd2e208561dda6a3.
• Refreshed TensorRT-LLM/TokenSpeed head anchors and profiler/benchmark references to 2026-06-27 heads.
• Added tools/check_open_pr_watch.py and generated model-pr-optimization-history/open-pr-watch.md; the tool now falls back to anonymous REST when gh api is rate-limited and fails closed if every repo fetch fails.
• Added root update_prompt.md for repeating this whole workflow.
• Fixed cookbook validator behavior so missing-command help captures do not look like unsupported framework flags.

B200 validation artifacts:

• Root: /data/bbuf/ai_infra_skills_pr72_20260627
• 5 SGLang serving smokes: Qwen2.5 0.5B/1.5B/3B/7B and Qwen3-8B, one B200, each cleaned after run.
• MiniMax-M3 container smoke logs under minimax_m3/.
• SGLang profiler live capture: /data/bbuf/ai_infra_skills_pr72_20260627/profiler/qwen25_0_5b_live
• NCU smoke: /data/bbuf/ai_infra_skills_pr72_20260627/profiler/ncu_smoke_qwen_matmul.ncu-rep

Validation:

• Local python3 -m pytest -q: 116 passed.
• Local pre-commit run --all-files --show-diff-on-failure: passed.
• Remote B200 subset: 17 passed, 160 subtests passed.
• Remote cookbook validator with captured help snapshots: passed.
• PR lint check is passing on commit a0ca475.