azcosmos: only mark region unavailable on NotSent network errors

[tvaron3]

Summary

Changes attemptRetryOnNetworkError in sdk/data/azcosmos/cosmos_client_retry_policy.go so that MarkEndpointUnavailableForRead/ForWrite is only invoked when the transport error is classified as connectionErrorNotSent — meaning we are sure the request never reached the service (DNS failure, TCP connect refused/unreachable, TLS handshake failure, etc.).

For ambiguous transport errors (EOF, RST, transport-level timeout mid-exchange — the request may have been processed server-side), the region is no longer marked unavailable. A single mid-exchange failure is too weak a signal to declare a regional outage for all concurrent and future traffic on the client.

The 403 endpoint-failure path (attemptRetryOnEndpointFailure) is intentionally unchanged — that path is a service response, not a transport error.

Behavior change

Scenario	Before	After
Read, NotSent	Mark read, demote-to-tail, failover via resolveFromHead	Unchanged
Read, ambiguous	Mark read, demote-to-tail, failover via resolveFromHead	No mark. Failover by bumping retryContext.retryCount so the next iteration of the Do loop resolves a different locationIndex. Mirrors how the 503/408 paths already failover without marking.
Write (multi-master), NotSent	Mark read+write, demote, failover	Unchanged
Write (multi-master), ambiguous	Mark read+write, no retry	No mark, no retry
Write (single-master), NotSent	Mark read only, no retry	Unchanged
Write (single-master), ambiguous	Mark read only, no retry	No mark, no retry
403 WriteForbidden / DatabaseAccountNotFound	Mark + forced async refresh + failover	Unchanged

Tests

• TestAmbiguousConnectionErrorReadFailsOver — rewritten to use the two-server routingMockTransport pattern (same as TestConnectionErrorReadFailsOverWhenGlobalEndpointIsUnreachable). Asserts the ambiguous-read retry actually reaches the second region (goodSrv.Requests() == 1), retryCount == 1, and the location-cache unavailability map stays empty.
• TestAmbiguousWriteMarksEndpointUnavailableForRead → renamed TestAmbiguousWriteDoesNotMarkEndpointUnavailable and inverted to assert the location cache stays empty.
• All NotSent-path tests (TestNotSentConnectionErrorMultiMasterWriteFailsOver, TestConnectionErrorGivesUpAfterSingleCrossRegionFailover, TestDnsErrorRetry, TestConnectionErrorReadFailsOverWhenGlobalEndpointIsUnreachable, etc.) pass unchanged.

Notes

• Comments in attemptRetryOnNetworkError were also significantly shortened.
• The change is internal to the retry policy; no public API change.

Co-authored-by: Copilot 223556219+Copilot@users.noreply.github.com

[tvaron3]

Superseded by upstream PR Azure#26915