ecosystem: publish suspect-zones benchmark corpus as a HuggingFace dataset

[devswha]

Why it matters

README headlines (91% / 76% / 13–25%) are unverifiable without the corpus. Publishing as HF dataset is the NLP-research reproducibility move and the prerequisite for a citation / arXiv paper.

Acceptance criteria

• huggingface.co/datasets/devswha/patina-suspect-zones with current tests/fixtures/suspect-zones/ contents
• Data card: per-source licensing, language coverage, how 91/76 were computed, known biases
• CI step (manual trigger ok for v1) re-uploads on tagged release
• Cross-linked from README, FAQ, research notes, latest.md
• Licensing gate: review each fixture for redistribution rights before upload

[devswha]

Wave 4 triage: this stays open.

Concrete blocker: a HuggingFace dataset publication needs corpus governance first: licensing/PII audit, dataset-card contents, split/schema decisions, and CI/publish ownership. The campaign did not have a vetted publishable corpus to upload.

[devswha]

Refreshed the publication bundle from main bb905f67a8b10c7e778043b791b823cace9b80df (8.3.1). Export and the default offline publisher validation both passed for all 49 reviewed public fixtures. The bundle includes the dataset card, MIT license, test JSONL and source manifest; private rebaseline/panel data remain excluded.

No upload occurred. The target is still devswha/patina-suspect-zones; publication and public cross-links remain pending target-namespace access and successful remote verification. An exported and validated bundle does not complete this issue.