[P1] ML.NET workflow intent classifier — trained model upgrade path

[JerrettDavis]

Objective

Replace/augment the TF-IDF heuristic classifier (#401) with a trained ML.NET binary classifier for higher accuracy workflow intent detection.

Approach

• ML.NET SdcaLogisticRegression or LbfgsMaximumEntropy
• Train on synthesized prompt pairs (workflow vs conversation)
• Ship as .zip model file, loaded at Gateway startup
• <1ms inference, zero external deps

Prerequisites

• [P0] TF-IDF workflow intent classifier — pre-agent prompt routing #401 (TF-IDF classifier) proves out the concept and routing pipeline
• Training data synthesis: generate 500+ labeled prompt pairs from real agent sessions + synthetic augmentation

Deliverables

• Training data generation script (pull from agent session history + augment)
• ML.NET training pipeline
• MlNetIntentClassifier : IPromptIntentClassifier implementation
• A/B comparison: TF-IDF vs ML.NET accuracy on holdout set
• Hot-swap support: replace classifier at runtime without restart

Blocking Order

Depends on #401 (TF-IDF proves concept)