Background

Issue #2190 introduced Candle-backed injection classifier infrastructure. The FeedbackDetector in zeph-core currently supports two modes:

• Regex (default)
• Judge (LLM-backed)

A third Model variant using the ClassifierBackend trait would enable fast, offline, CPU-based feedback detection without an LLM round-trip.

Required work

1. Add DetectorMode::Model { repo_id: String } variant to the config enum in zeph-config
2. Add ClassifierBackend integration to FeedbackDetector in zeph-core
3. Add zeph-core dependency on zeph-llm classifier module (or extract ClassifierBackend trait to a shared crate)
4. Wire detector_mode = "model" in config parsing
5. Add tests for the new variant

Notes

• FeedbackDetector lives in crates/zeph-core/src/agent/feedback_detector.rs
• ClassifierBackend trait is in crates/zeph-llm/src/classifier/mod.rs
• This is a cross-crate dependency change — requires careful consideration of the dependency graph
• Deferred from PR test(candle): add integration tests for Candle-backed classifier models #2190 (test infrastructure PR) as it requires implementing a new feature, not just tests