Point the main Cosmos live-test stage and the Http2 stage at fixed,
self-owned Cosmos accounts (RG sdk-ci) instead of provisioning an account
per pipeline run. This removes the central EngSys service-connection /
on-the-fly-provisioning dependency for the key-based tests.

- Add sdk/cosmos/pipeline/ with a single-secret JSON mechanism:
  - live-test-accounts.schema.json + sample
  - resolve-cosmos-test-account.sh (bash+jq) resolves one account from the
    single secret (sub-config-cosmos-azure-cloud-test-resources) by selector
    and exports ACCOUNT_HOST/ACCOUNT_KEY(+SECONDARY) using the azure-sdk
    double-set convention so the key is masked in logs yet reaches the test env
  - resolve-cosmos-test-account.tests.sh (local tests)
  - resolve-test-account-steps.yml (reusable PreTestRunSteps template)
- tests.yml: Cosmos_live_test + Cosmos_Live_Test_Http2 set
  DisableAzureResourceCreation:true and run the resolve pre-step.
- live-platform-matrix.json: add per-leg AccountSelector.

Thin-client, GSI, Kafka and Spring stages are intentionally left on the
existing provisioning path (they need feature-enabled accounts or a separate
AAD pipeline) and will be migrated in follow-ups.

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

…test-sub-tenant

# Conflicts:
#	sdk/cosmos/live-platform-matrix.json

…count-safe

Follow-up to the fixed-account live-test wiring: bring the account
provisioning tooling into the repo and fix tests that assumed each run
gets its own freshly-provisioned account.

- Add sdk/cosmos/pipeline/account-provisioning/ (moved from the
  cosmos-sdk-copilot-toolkit repo):
  - New-CosmosLiveTestAccounts.ps1 creates the fixed sdk-ci accounts and
    outputs the cosmos-live-test-accounts JSON; reconciles capabilities on
    existing accounts (EnableNoSQLVectorSearch) via ARM PATCH.
  - Set-CosmosLiveTestAccountsSecret.ps1 publishes that JSON to the secret.
  - cosmos-live-test-accounts.definition.json defines the accounts; the 5
    general-purpose accounts enable EnableNoSQLVectorSearch to match the
    ARM test-resources.json capability set.
- ReadFeedDatabasesTest: readAllDatabases is account-global; on shared
  fixed accounts other runs create/delete DBs concurrently. Assert only
  that the databases created by this test are present (containment) via a
  new FeedResponseListValidator.containsResourceIds, instead of asserting
  the exact account-wide set/count/page-count.

Verified locally: ReadFeedDatabasesTest failed 3/3 under concurrent DB
churn before and passes 3/3 after; NonStreamingOrderByQueryVectorSearchTest
passes 4/4 once EnableNoSQLVectorSearch is enabled on the account.

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

…/pipeline

- Remove Set-CosmosLiveTestAccountsSecret.ps1 and its references; the Key
  Vault / ADO variable secret is now updated manually with the JSON that
  New-CosmosLiveTestAccounts.ps1 outputs. Docs updated accordingly (removed
  the KV-publish usage, the dated "Verified run" note, and the macOS auth
  note).
- Move the cspell allow-list for pipeline domain/code tokens (sdkci, pmerge,
  Pfast, issecret, esac, WHATIF) out of the root .vscode/cspell.json into a
  nested sdk/cosmos/pipeline/cspell.json, which cspell merges automatically.
  Keeps the change scoped to files under sdk/cosmos.

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

readOffers is an account-global read. With live tests running against
shared fixed accounts, concurrent legs create/delete containers (and
their offers), so asserting the exact account-wide offer set/count/page
count is non-deterministic. Assert containment of only the offers for
containers created by this test instead.

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

… guards

- Align fixed-account regions with the ARM bicep (test-resources.json):
  regionDefaults -> Central US (single), Central US + East US 2 (multi).
  This fixes the fault-injection availability-strategy tests whose region
  assertions require the preferred region (East US 2) to exist on the account.
  Rewire New-CosmosLiveTestAccounts.ps1 to consume regionDefaults as the single
  source of truth. Remove a stale gsi-single-session preferredLocations landmine.

- Isolate the partition-split-forcing profiles (-Pcfp-split, -Psplit) onto
  dedicated accounts (single-session-split, single-strong-split,
  multimaster-multiregion-session-split) so their split churn no longer
  contends with the shared query/fast/direct legs on the same account.

- Harden ReadFeedOffersTest and ReadFeedDatabasesTest so the containment
  assertion cannot silently degrade to a no-op on an empty expected set.

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

Live tests now share fixed accounts across concurrent CI legs, so
database/container CRUD can hit control-plane throttling (metadata 429 /
substatus 3200) and 408 timeouts that the default client retry (9/30s)
does not absorb. Add TestSuiteBase.executeControlPlaneWithRetry, which
retries only transient control-plane failures (408/429/500/503) and
rethrows everything else immediately so negative-path assertions (400/409)
are preserved. Wrap the createContainer calls in
MultiMasterConflictResolutionTest and the createDatabaseIfNotExists call
in CosmosTracerTest (with a tracer reset between attempts so a throttled
attempt's spans do not leak into the trace assertions).

The default SDK client throttling behavior is intentionally left unchanged
so tests that assert on retry counts (MaxRetryCountTests, the
fault-injection availability-strategy tests) are unaffected.

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

These change-feed tests are known-flaky (see Azure#48064, Azure#45216); running on
shared fixed accounts surfaces transient 408/429/503 more often during
forced partition splits and change-feed setup.

- Broaden SplitTestsRetryAnalyzer to retry not only SplitTimeoutException
  but also a transient CosmosException (408/429/503) anywhere in the cause
  chain, so verifyLeasesOnRestart_AfterSplit tolerates the 503s a split
  legitimately produces.
- Add FlakyTestRetryAnalyzer to readFeedDocumentsStartFromCustomDateForMultiWrite_test
  (pkversion + epkversion) and ChangeFeedTest.changesFromPartitionKey_AfterInsertingNewDocuments,
  matching the precedent set for their sibling CFP tests in Azure#48064.

No production SDK code or default client behavior is changed.

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