Add CI artifacts for sanitized benchmark and profile outputs

[tony]

Summary

Keep CI artifact upload separate from the local observability work on streamline-02. The local profiler and benchmarker should land first; this issue can add CI upload support afterward without mixing workflow design into the profiler branch.

CI runners do not have representative local agent-history stores, so uploaded artifacts should be treated as harness-smoke evidence unless they are generated from explicit sanitized fixtures. Real bottleneck evidence should still come from local profiles.

References

• actions/upload-artifact at v7.0.1: https://github.com/actions/upload-artifact/tree/v7.0.1
• upload action metadata: https://github.com/actions/upload-artifact/blob/v7.0.1/action.yml

Proposed work

• Add a dedicated workflow or manual workflow dispatch for sanitized benchmark/profile artifacts.
• Consider a small reusable composite action in this repository if multiple jobs need the same upload policy.
• Use fixture-only profiler or benchmark payloads in CI; do not read real user agent stores.
• Upload JSON/Markdown artifacts with explicit names that include the component and GitHub run metadata.
• Use if-no-files-found: error for required artifacts and keep include-hidden-files: false unless there is a reviewed reason to change it.
• Set a short retention period appropriate for diagnostic artifacts.

Non-goals

• Do not treat CI timings as representative bottleneck evidence.
• Do not upload prompt text, conversation text, raw argv, local absolute paths, secrets, or hidden files.
• Do not make this part of the streamline-02 local profiler branch.

Acceptance criteria

• CI can upload sanitized profiler/benchmark artifacts from fixture-backed runs.
• Artifact names and descriptions make limits and components visible.
• The workflow documents that local profiles remain the source of real bottleneck evidence.
• Any reusable action is small, repo-local, and wraps actions/upload-artifact rather than reimplementing upload behavior.