[research milestone] Build an externally labeled holdout benchmark

[niansia]

ContextSec reports authored-regression and frozen holdout results honestly; the next evidence milestone should measure unseen repositories without implementer-known answers.

Acceptance criteria:

• Declare a repository sampling frame and inclusion/exclusion rules.
• Freeze commit-pinned repositories before ContextSec profiling.
• Collect annotations from reviewers who did not implement the relevant detectors.
• Preserve reviewer disagreement instead of silently reconciling it.
• Report precision/recall by pack and framework, with unsupported-stack results separate.
• Never install, import, build, test, or execute target code.
• Publish the dataset license/provenance and retain the research-preview claim boundary.