resolveTimeInActive decompressed every fsynced block from the front of the active segment inside the websocket handshake. a time cursor older than the hot tail but not yet sealed - a band hours wide - took 40s+ to resolve mid-fill and timed the connect out entirely: coral retried the same cursor against the same host forever (its zat 0.3.25 has no idle watchdog and never resets backoff). block boundaries now come from the 8-byte length prefixes alone and only log2(blocks) probes decompress, matching how sealed segments already resolve via their block index. regression test sweeps every microsecond boundary across a ten-block active region against the exhaustive definition. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>