De-flake various tests - #8703

Claude finished @MauriceVanVeen's task in 1m 20s —— View job

Review of PR #8703 — De-flake various tests

• Fetch PR diff and description
• Review each changed file for correctness, concurrency, security
• Check for inline issues
• Post summary review

Summary: This PR fixes test flakiness plus one small production bug in server/jetstream_cluster.go/server/stream.go.

Production change (matchSourceIndexNames)

• server/jetstream_cluster.go:6958 now calls the newly extracted matchSourceIndexNames(sa.Config, osa.Config) under js.mu.Lock()/Unlock() before the reflect.DeepEqual short-circuit check in processClusterCreateStream. This mirrors the existing logic already used in jsClusteredStreamRequest (which holds js.mu for its whole body), so the locking is consistent and not nested/double-locked — I checked the surrounding RLock/Lock blocks and they're sequential, not nested.
• Root cause this addresses: previously the index name (iname) on sourced streams wasn't reconciled before the DeepEqual check here, so an idempotent AddStream could fail to short-circuit after a leader change (since iname isn't part of the wire-encoded config), causing the reported timeout. The fix looks correct and properly scoped.

Test changes

• TestFileStoreAccessTimeSpinUp (filestore_test.go): switches from a flat NumGoroutine() comparison to counting only ats.Register.func* goroutines via stack dump, correctly accounting for the fact that ats registration is a process-wide singleton shared across parallel tests. Confirmed the goroutine name substring matches the actual closure in server/ats/ats.go.
• TestJetStreamClusteredStreamCreateIdempotentWithSources (jetstream_cluster_3_test.go): reworked to step leadership through every server (not just directly to the meta leader) and retry step-downs since leadership transfer is best-effort; also asserts iname is populated on all servers before the idempotent-create assertion. Logic reads correctly, using errNotLeader (defined in raft.go) as an expected/retryable error.
• TestJetStreamClusterStreamAdvisories, TestJetStreamAllowMsgCounter, TestNoRaceJetStreamFileStoreLargeKVAccessTiming, TestNoRaceSeqSetRelativeSpeed: all reasonable retry/threshold relaxations for timing- and propagation-sensitive assertions. The AVL test's 3x→5x threshold loosening slightly weakens the performance regression check but is a defensible tradeoff for CI stability.

No correctness, concurrency, or security issues found. No inline comments posted.