Background

Issue #2190 requires integration tests for iiiorg/piiranha-v1-detect-personal-information (NER/PII detection model) with token-level label output. This is distinct from the sequence classification backend already implemented in PR #2198.

Why this is deferred from #2190 test PR

The current CandleClassifier uses DebertaV2SeqClassificationModel — a sequence-level classifier that produces one label per input. piiranha is a token-level NER model that requires DebertaV2NERModel (available in candle-transformers 0.9.2), which produces per-token logits with shape [batch, seq_len, num_labels].

Testing piiranha with the existing CandleClassifier is not possible — the model architectures are fundamentally different. Adding NER support requires:

1. A new backend struct (e.g., CandleNerClassifier) wrapping DebertaV2NERModel
2. A new trait or extended ClassificationResult carrying token-level spans
3. BIO/BIOES span decoding logic (B-EMAIL, I-EMAIL → span extraction)
4. New ClassifierBackend variant or a separate NerBackend trait

Required tests (once implemented)

• Load iiiorg/piiranha-v1-detect-personal-information via DebertaV2NERModel
• Token-level label output: verify EMAIL / PHONE / SSN entities are tagged correctly
• Multi-language test: one English PII phrase, one German PII phrase

Notes

• DebertaV2NERModel is already present in candle-transformers 0.9.2 (confirmed)
• Related: feat(classifiers): replace regex heuristics with Candle-backed lightweight classifiers #2185 (classifier infrastructure), feat(core): add DetectorMode::Model for classifier-backed feedback detection #2210 (FeedbackDetector Model variant), feat(classifiers): Candle-backed injection classifier infrastructure (#2185) #2198 (sequence classifier PR)