Native retracting OFFSET: preserve upstream stateful mini-batch changelog order

Remaining problem

Hidden-rank retracting OFFSET still falls back when an upstream stateful mini-batch operator can reorder or collapse its input changelog. Released Flink retains emitted row kinds inside the rank buffer, so those input-order differences can affect later full-row retraction matching and the materialized result.

The retained-row-kind implementation shipped in PR #132, merged as 40290ec9. It preserves sort-key counts independently of retained payloads, emits the per-record cascade, honors both update-before settings, and restores retained kinds/counts across memory and RocksDB. The admitted path passed exact ordered changelog comparisons, the full Java/Rust suites, unchanged upstream Rank/Limit/SortLimit tests, and all 24 required PR checks.

Current coverage and release measurements live in LIMIT and Top-N. Unique-key update-fast OFFSET and projected-rank paths remain covered separately.

Remaining scope

• Reproduce the upstream stateful mini-batch changelog order needed by hidden-rank retracting OFFSET, including grouped pipelines whose bundles collapse intermediate changes.
• Preserve full-row retraction matching, retained row kinds and independent sort-key counts through that composition.
• Keep the precise retracting OFFSET requires unchanged upstream mini-batch changelog order gate until the full pipeline is verified.

Acceptance

• Compare exact raw changelogs and final materializations against released Flink with stateful upstream mini-batches, duplicate payloads, tied keys, NULL ordering and group deletion.
• Cover both update-before settings and retractions after prior positional emission.
• Verify checkpoint continuation and both state backends for the admitted composition.
• Assert native execution of the full grouped/OFFSET pipeline and update the LIMIT/Top-N coverage pages when the remaining gate is removed.

Start with the LIMIT and Top-N coverage pages and the existing native retracting OFFSET gate. Reproduce the grouped stateful mini-batch composition, compare exact raw changelogs and final materializations with released Flink, then verify checkpoint continuation, both state backends, update-before settings, native execution, and the coverage-page updates.