fix(transactions): harden Azure Table transactional state storage
Azure Table transactional state storage mutates its cached key and prepared-state records before submitting a transaction, so a failed write could leave the storage instance unsafe to reuse. A paginated entity query can also cross a concurrent multi-batch write because continuation pages are independent requests, while conflict failures previously exposed limited context.
Require a successful reload before reuse after a failed mutating store, preserve strict ETag conflict detection, and emit payload-free operation diagnostics with the failed action details supplied by Azure Table Storage. Load queries the key and state rows together. The normal single-page case returns the snapshot in one storage operation. Every physical write batch atomically upserts ! and ~ boundary rows with the same fresh fence value; their keys sort before and after all transactional rows, so matching values from a completed paginated query prove that no boundary-aware batch committed while its pages were read. Torn boundary values retry up to a bound without an additional key read.
Legacy partitions without boundary rows retain the previous one-query behavior. During a rolling upgrade, older writers do not advance existing boundary rows, so paginated torn reads remain intentionally possible until every writer is upgraded; the stronger boundary guarantee applies once the cluster has fully transitioned.
Each physical batch still conditionally advances the key and now reserves two boundary slots, including the final partial batch. A fresh key carries valid empty prior metadata before any split prepare batch can persist it, so a failure after the first physical batch leaves a recoverable intermediate snapshot without prematurely publishing incoming commit metadata or sequence position. Deterministic coverage exercises one-query loads, stable and torn boundary pagination, legacy no-boundary acceptance, bounded churn, fresh split recovery, per-batch fence uniqueness, no-change conflicts, stale-writer recovery, failed-store reuse rejection, and diagnostic redaction.
Azure entity group transactions atomically mutate up to 100 entities sharing a partition, and the transaction batch protocol plus TableTransactionActionType expose mutation actions but no transactional read batch. Query Entities can return a single strongly consistent response, but larger or slower queries resume through subsequent requests under the documented query timeout and pagination semantics. Once all writers maintain the boundary rows, paginated reads become self-validating without another storage request.