Final Iceberg validation

The native suite passes 298 tests, with 13 batch-only exclusions (311 invocations across 35 classes). The stock leg passes the matching 141 streaming SQL/control invocations, with the same 13 exclusions. The verifier requires identical inventories for 14 upstream streaming SQL cases and 127 range/dynamic controls. Native coverage includes twelve captured-channel recovery/rescale fixtures, four mid-file source recoveries, SQL writes, deletes, reader predicates, schema evolution, task dispatch, quota failures and released-handle checks.

All six release Rust unit tests and 126 release+mimalloc Criterion fixture checks pass. Criterion smoke validates fixtures and allocation diagnostics; it is not a new timing baseline. summary.json records report inventories, commands and library hashes; reports.tar.gz retains sanitized XML, logs and allocation CSV. Java tests run serially with 4 GiB heaps and the same staged release native libraries as the Nexmark matrix.

The first full rerun exposed an early recovery-savepoint race and unclaimed Arrow batches left in the fetch queue after reader cancellation. The retained native-sql-recovery.pre-cleanup.log and recovery-readiness.log show those failures. The cancellation registry now owns queued batches until emission; recovery waits for every task to run. All thirteen targeted reader/recovery cases then pass, including zero Arrow/task reservations after each mid-file cancellation and restore, and a new test that keeps claimed batches valid across cancellation. The complete suite was rerun successfully after that fix. The unchanged released stock leg is retained from the first successful stock/control run.