Strengthened Input Analysis for Long and Mixed Content

Strengthened the input analysis foundation for handling long content that mixes natural language, code, and structured data.

Instead of applying heavy analysis uniformly across the entire input, the system now follows a staged approach: first observing lightweight signals and possible boundaries, then applying deeper analysis only where needed. This makes it easier to understand the structure of long inputs while keeping processing overhead under control.

The new approach forms the core of the Input Foundation and provides the basis for future chunking, tag and keyword extraction, routing, and other structured input processing.