FEAT Proposal: Turkish Prompt-Injection dataset loader (design/provenance discussion)

Opening this as a design-first proposal before asking for a merge, following the catalog-direction guidance from @adrian-gavrila on #2175 (propose and get maintainer direction before building).

What

A SeedDatasetProvider loader for the open Turkish Prompt-Injection dataset (3nesdeniz/turkish-prompt-injection-1k, CC BY 4.0): paired benign/attack Turkish examples across instruction override & system-prompt extraction, jailbreak personas, obfuscation & data exfiltration, and agentic tool abuse, with Turkish morphological variation.

Why it could be valuable

• Multilingual coverage gap: PyRIT's injection-related seed datasets are English-first. Guardrails that pass in English often fail to generalize; Turkish (rich morphology) is a useful stress test for OWASP LLM01.
• False-positive measurement: the benign hard negatives share surface vocabulary with the attacks, so detector over-blocking is measurable, not just recall.

Provenance (being upfront)

This is not backed by a peer-reviewed paper today. What exists: an openly published, CC BY 4.0 dataset with a CITATION.cff, a pinned immutable HF revision, and a documented deterministic/reproducible generation methodology; split-isolated train/val/test with no template leakage.

Question for maintainers

Given that most loaders are paper-backed, what evidence-of-value / provenance bar would you want here before this is a candidate — e.g. a peer-reviewed or preprint write-up of the methodology, third-party usage, or something else? I'd rather align on direction first than repeat the sequencing issue from #2174/#2175.

A tested implementation is ready as a draft PR (#2365) purely for reference; I'm not requesting review of it until the direction here is settled.