Production feedback loop management for ML systems in the Crucible ecosystem.
This library provides telemetry ingestion, quality assessment, drift detection,
data curation, retraining triggers, and export capabilities for Elixir-based
machine learning pipelines.

Core Features:

Ingestion Pipeline
  - Batch-buffered event ingestion with configurable flush intervals
  - Automatic PII sanitization for emails, phones, SSNs, credit cards
  - User ID hashing for privacy protection
  - Ecto schema validation for incoming events
  - Telemetry events for monitoring ingestion health

User Signal Capture
  - Support for thumbs up/down, regenerate, edit, copy, share, report signals
  - User edit tracking with edited response storage
  - Signal scoring utilities for downstream curation

Quality Assessment
  - Format detection for JSON, markdown, and plain text responses
  - Length validation with configurable bounds
  - Refusal pattern detection for unwanted model behaviors
  - Repetition and stuttering detection via trigram analysis
  - Optional LLM-as-judge integration point
  - Rolling average quality score computation

Drift Detection
  - Statistical drift via Kolmogorov-Smirnov test and PSI metrics
  - Embedding-based drift using centroid distance comparison
  - Output drift tracking for quality and response length shifts
  - Pluggable embedding client behavior for provider flexibility

Data Curation
  - High-quality example selection based on quality scores and signals
  - Hard example mining for low-quality or negatively signaled events
  - Diversity selection via embedding space distance from centroid
  - User edit example prioritization for preference learning
  - Automatic deduplication across curation strategies

Retraining Triggers
  - Drift threshold trigger when distribution shift exceeds limits
  - Quality drop trigger when rolling average falls below threshold
  - Data count trigger when sufficient examples are collected
  - Schedule trigger for time-based retraining intervals

Export Capabilities
  - JSONL export for standard training pipelines
  - HuggingFace dataset directory format with metadata
  - Parquet placeholder for future columnar export
  - Preference pair export from user edits for DPO training
  - Export batch tracking with idempotent marking

Storage Backends
  - Ecto/Postgres backend for production deployments
  - In-memory backend for testing and local development
  - ClickHouse placeholder for high-volume telemetry
  - Storage behaviour for custom backend implementation

Crucible Integration
  - ExportFeedback stage for pipeline-based data export
  - CheckTriggers stage for automated retraining decisions
  - Context artifact storage for inter-stage communication

Infrastructure
  - Comprehensive test suite with Mox-based mocking
  - ExMachina factories for test data generation
  - Telemetry instrumentation throughout the pipeline
  - MIT license for open source distribution