Cosmos: Harden container recreation tests

### Feature Summary

Expand container recreation coverage to verify both safe recovery and deterministic rejection when a name-addressed container is deleted and recreated with a different definition.

### Feature Description

PR #5219 implemented one-shot recovery for name-addressed container operations after a verified stale-generation response. Existing end-to-end coverage recreates the container with the same `/pk` partition-key path, leaving important generation-boundary behavior unverified.

Add focused tests that recreate a container under the same name with a new RID, a different partition-key path, and changed throughput. Cover at least:

- compatible point operations recovering once using refreshed metadata, routing, and session state;
- operations built with the old partition-key shape failing deterministically rather than reaching the replacement container;
- queries and transactional batches across the generation boundary;
- continuation tokens, explicit session tokens, stale EPK ranges, and other state that must not cross generations;
- throughput reads and replacements after recreation;
- the standard Gateway and appropriate emulator/live-test legs.

Acceptance criteria:

- tests demonstrate that safe operations recover at most once;
- generation-bound state is never silently retargeted;
- changed partition-key and throughput scenarios have positive and negative coverage;
- failures assert stable status/substatus behavior where the contract is defined;
- Gateway 2.0-specific work remains out of scope and is tracked separately.

### Use Case

Container delete-and-recreate workflows are rare but correctness-critical: stale clients must never send an operation intended for the old physical container to the replacement container.

### Alternatives

Rely on same-schema recreation tests and unit coverage. This does not validate the highest-risk changed-schema and stale-state scenarios end to end.

### Additional Context

Sub-issue of #3526. Related implementation and tracking: #4427, #5197, and PR #5219.

Start by reading PR #5219 and the existing end-to-end container recreation coverage, then compare the standard Gateway with the emulator/live-test legs. Add focused cases for changed RID, partition-key, throughput, query, batch, token, session, and stale-range behavior; done means safe recovery happens at most once and incompatible state fails with the defined status and substatus.

Written by the indexing model from the issue text.