[fxbing]

Purpose

Closes #4909

Fix KV batch scans repeatedly targeting an offline cached leader when callers recreate the scanner using the same Connection and Table.

Brief change log

• Invalidate table or partition bucket metadata on InvalidMetadataException, including NetworkException, and on expired or unknown scanner sessions.
• Cover synchronous dispatch failures, asynchronous failures, and error responses for open and continuation requests.
• Keep the failing scan terminal and preserve existing TOO_MANY_SCANNERS retries and response-buffer cleanup. Recovery happens through a caller-created scanner.

Tests

• Added a three-node regression: cache leader A, stop A, wait for the new leader, verify the first scan fails, then rebuild the scanner once on the original Connection and Table and read all eight rows. The rebuilt scan fails before the fix and succeeds afterward.
• Added coverage for error classification, open and continuation failures, partition isolation, and preservation of the original failure when invalidation fails.
• Passed 84 tests across KvBatchScannerTest, TableKvScanITCase, BatchScannerITCase, KvSnapshotAndLogBatchScannerTest, CompositeBatchScannerTest, and MetadataUpdaterTest. Checkstyle, Spotless, and RAT passed.

Full repository verification and the JDK 8 CI matrix were not run locally.

API and Format

No public API, configuration, RPC protocol, or storage-format changes.

Documentation

No documentation changes.

Generative AI disclosure

Generated-by: OpenAI Codex

Implementation assisted by OpenAI Codex; independently reviewed with DeepSeek.