Total bytes read from the source (≈ file size + re-reads on growth).
Largest single top-level record observed, in bytes.
Target window size in bytes (the pool).
Number of times the window slid forward.
Physical window buffer size actually used (grows past pool only if a
single record exceeds the pool).
Diagnostics from a streaming index build — how hard the moving window had to work, so a pool-size sweep can compare cost against a resident parse.