Build a labeled baseline dataset of interesting real pipeline failures

Part of the Pipeline Analysis · Azure SDK Tools Agent internship project. - **Area:** Evaluation - **Sequence:** 9 - **Estimate:** 5 implementation days - **Assignee:** @ReilleyMilne ## Depends on: - #16195 (Build a standardized artifact discovery and retrieval layer for pipeline runs) - #16196 (Add shared regression fixtures from real pipeline failures) ## Definition of done A scripted collection process identifies representative real failures, gathers the needed artifacts and metadata, and produces a curated dataset with expected diagnoses and scoring criteria, using approved stored inputs where needed for evaluation.