[WIP][experimental] add agentic trace replay benchmark infrastructure

[cquil11]

Trace replay benchmarking for agentic coding workloads using real Claude Code traces. Includes:

• Trace replay scripts for H200, MI355X, B200 (vLLM-based)
• kv-cache-tester submodule (trace replayer + 522 anonymized traces)
• AIPerf submodule (alternative synthetic benchmarking)
• Pareto frontier plotting and sweep aggregation
• Metrics collector (prometheus scraper + visualization)
• Workload distribution analysis
• GitHub Actions workflow with per-TP sweep configs
• MI355X runner SCRIPT_SUFFIX support

[github-actions]

Thanks for the contribution! For vLLM & SGLang, please ensure that your recipes is similar to the official vLLM recipes and/or the SGLang cookbook

• https://cookbook.sglang.io/
• https://docs.vllm.ai/projects/recipes/en/latest/

If it is not, please create a PR first before we can merge your PR into the master branch. Let's ensure that the documentation is first class such that the entire ML community can benefit from your hard work! Thank you

• https://github.com/vllm-project/recipes
• https://github.com/sgl-project/sgl-cookbook

In general, fix this by explicitly setting a permissions block either at the workflow root (to cover all jobs) or on individual jobs, granting only the minimal scopes required. Since this workflow does not rely on GITHUB_TOKEN for write operations (it uses secrets.REPO_PAT for checkout and only handles artifacts), we can safely set contents: read at the workflow level. This documents intent and ensures GITHUB_TOKEN cannot be used for repo writes, even if org defaults are permissive.

The best minimal fix without changing existing behavior is: add a permissions: block at the top level, right after the existing name/run-name and before on:. Set contents: read as the default for all jobs. No additional imports, methods, or definitions are needed because this is purely a YAML configuration change. The jobs (generate-matrix, sweep, collect) will automatically inherit these restricted permissions unless they define their own permissions (which they currently do not).

Concretely:

• Edit .github/workflows/multiturn-sweep.yml.
• Insert:

permissions:
  contents: read

between the run-name: line and the on: block (after line 2 and before line 4 in the provided snippet). This satisfies CodeQL’s requirement and implements least-privilege defaults for GITHUB_TOKEN across the workflow.

In general, the fix is to declare an explicit permissions: block applying the least privileges required. This can be done at the workflow root (applies to all jobs) or per job. Since none of the shown jobs needs to write to the repo or to other resources via GITHUB_TOKEN, we can safely restrict it to read-only repository contents. The minimal recommended setting is permissions: contents: read at the workflow root.

Concretely, in .github/workflows/multiturn-sweep.yml, add a top-level permissions: block right after the run-name: (or after on:) so it applies to all jobs (generate-matrix, sweep, and collect). Set it to:

permissions:
  contents: read

This change does not alter existing functionality: the actions/checkout steps use an explicit PAT via token: ${{ secrets.REPO_PAT }}, and all other actions (download/upload-artifact, setup-python, etc.) work with read-only contents permissions on GITHUB_TOKEN. No additional imports, methods, or other definitions are needed.