Context

Issue #2185 proposes replacing regex heuristics with Candle-backed lightweight classifiers. The research (CI-178/CI-180) confirmed:

• candle_transformers::models::deberta_v2 supports both sequence classification and NER
• EmbedModel in candle_provider/embed.rs already covers 90% of the boilerplate
• parry-guard (vaporif/parry-guard) proves the pipeline works in Rust today

Required Tests

1. ClassifierModel unit tests (once implemented per #2185)

• Load protectai/deberta-v3-small-prompt-injection-v2 via Candle deberta_v2 module
• Score a known injection string (should exceed threshold)
• Score a benign string (should be below threshold)
• Verify tokenizer handles SentencePiece correctly

2. PII/NER model tests

• Load iiiorg/piiranha-v1-detect-personal-information via Candle deberta_v2 NER mode
• Token-level label output: verify EMAIL/PHONE/SSN entities are tagged correctly
• Multi-language test: one English, one German PII phrase

3. Performance benchmarks (criterion)

• Candle vs ort path latency for injection detection
• Cold start (model load) vs warm inference latency
• CPU-only target (no Metal/CUDA for CI)

4. Integration with FeedbackDetector

• detector_mode = "model" routes through zero-shot provider path
• Fallback to regex on provider timeout

5. Regression tests for #2189

• Verify cargo nextest run -p zeph-llm --features candle --lib compiles and passes after bug(candle): --features candle alone fails to compile (MessageMetadata unresolved) #2189 fix

Notes

• Tests requiring model downloads (#[ignore]) — run manually or in a separate CI job with HF cache
• CPU-only inference must complete in < 500ms per call in CI
• Future: many specialized lightweight models per task (strategic direction CI-178)