Summary

This PR improves robustness, reproducibility, and error handling across several parts of the TRACER pipeline.

Changes

Robustness Improvements

• Added a guard in src/tracer/analysis/qualitative.py to prevent failures when processing empty trace datasets.
• Improved small-dataset handling in src/tracer/fit/pipeline.py by prioritizing validation splits when calibration splits are not feasible.
• Added safe JSON serialization in src/tracer/policy/artifacts.py to prevent failures caused by non-serializable objects during manifest generation.

Consistency & Reproducibility

• Fixed embedding index behavior so requested similarity metrics are preserved consistently when FAISS is unavailable.
• Persist similarity metric configuration to disk and restore it during reload, ensuring identical behavior across environments and sessions.

API & Runtime Improvements

• Added embedding dimension validation in update() to provide clearer errors when new traces are incompatible with an existing model.
• Extended /predict_batch responses to expose additional metadata (preds and stage_id) for consistency with the local Python API.

Testing

Added dedicated tests covering:

• Empty dataset handling
• Small dataset split logic
• Embedding metric persistence
• Embedding dimension mismatch validation
• Index fallback behavior

Validation

• Executed the existing demo workflow after applying the changes.
• Verified that the new test cases pass successfully.
• Confirmed expected behavior for the affected edge cases.