test(server): remove the unlocked log-capture pattern from the remaining cmd/server tests (#301 follow-up) #310

Follow-up to #301 / #307. #307 fixed the harness-only race in TestHashMigrate_LogsMaxWriteLockHold_215: an unlocked read of the log-capture bytes.Buffer while background goroutines were still logging. The same anti-pattern remains in other cmd/server tests. The review of #307 also left four nits.

Same unlocked-capture pattern in 7 test files

These test files read a plain bytes.Buffer while the standard logger still points at it:

• cdn_detection_test.go (two sites, plus one writerFunc)
• issue199_node_not_found_test.go (two sites)
• memlimit_test.go
• neighbor_graph_geo_test.go
• nodes_pagination_postfilter_test.go
• ping_score_history_production_test.go
• schema_degradation_per_store_test.go

The hazard is not limited to the index builders. Several other goroutines can outlive a test and write into a later test's capture:

• OpenDB starts go d.healSchemaFlags().
• runDistanceIndexBuild logs every build.
• (*DB).Close logs the WAL checkpoint.
• Several recomputers log from their own goroutines.

Two fixes would close the class:

• Restore the writer before reading the buffer. log.SetOutput and (*Logger).output share outMu, so restoring first orders every in-flight logger write before the read. This is the cheapest option.
• Use one shared locked capture helper, like syncBuffer from test(server): data race between background index builds and the hash-migrate log capture (#301) #307.

Nits from the #307 review

1. neighbor_graph_geo_test.go:20-22. The comment claims a plain bytes.Buffer capture is safe under concurrent log.Printf. That is the reasoning test(server): data race between startBackgroundIndexBuilds and PacketStore.Load under -race (TestHashMigrate_LogsMaxWriteLockHold_215) #301 disproved: the logger orders writers against each other, but not against the harness's unlocked String(). Correct it.
2. hash_migrate_log_race_301_test.go:41. The fixed time.Sleep(200 * time.Millisecond) is what makes the test a reliable race detector. Consider polling on syncBuffer's written channel, or a read func() string handed to fn.
3. hash_migrate_log_race_301_test.go:28-31. If store.Load() fails, t.Fatal fires while the gate goroutine is still blocked on <-release. Add defer close(release) after installing the gate.

Acceptance

• No test in cmd/server reads a log-capture buffer without one of the two fixes above.
• go test -race -count=20 on the touched tests is green.
• No production code change.