Module compression_predictor 

Performance-floor compression ratio predictor (#2460).

A lightweight linear regression model that predicts compaction probe quality at a given compression ratio. Used to select the most aggressive compression ratio that keeps the predicted probe score above hard_fail_threshold.

Design

• No external ML crate dependencies — pure f32 arithmetic
• 4 input features: compression_ratio, message_count, avg_message_length, tool_output_fraction
• MSE loss with mini-batch gradient descent (continuous score target, not binary)
• Sigmoid output activation to bound predictions in [0.0, 1.0]
• Persisted as JSON via the compression_predictor_weights SQLite table
• Falls back to None (use default behavior) during cold start
• Training data sliding window: only the most recent N samples are retained

Structs

CompressionFeatures

Input features for the compression quality predictor.

CompressionModelWeights

Persisted model state (JSON-serializable).

CompressionPredictor

Compression quality predictor using linear regression with sigmoid output.