Problem

Full statistics walks can read every leaf and inspect every cell, making page diagnostics expensive on large collections. Sampling must happen before leaf reads to reduce I/O.

Proposed change

Add an opt-in sampling probability for the new row-store leaf diagnostics. Select eligible leaf references independently before loading them, with the same selection rule for cached and nonresident pages. Keep full walks as the default.

The existing custom-skip mechanism is a candidate. Internal structure still needs traversal, so sampling reduces expected leaf I/O without guaranteeing a fixed runtime.

Report sampling probability, references considered, pages selected, pages analysed, diagnostic exclusions and raw sample aggregates. Distinguish unselected pages from selected pages that cannot be analysed. Byte proportions must use totals from the same analysed sample.

Keep sampled results separate from full-walk gauges, including values returned by fast statistics. Initially exclude whole-tree extrapolation and confidence intervals; a small sample can miss rare large pages.

Acceptance criteria

• Tests demonstrate fewer leaf reads and analysed cells, with selection across the tree independent of cache residency.
• Selecting all pages matches a full walk on an unchanged tree.
• Test reproducible selection, empty samples and excluded pages.
• Preserve concurrent tree-walk safety and existing statistics behavior.
• Benchmark leaf reads, runtime and cache impact at several sampling rates.

Related: WT-18679, WT-18680.