test

  * Fix cluster retrieve blocking HTTP requests: fetchFromCluster now

    uses fire-and-forget async retrieve instead of synchronous 5s wait.

    Result is picked up on the next prefetch() call — no request ever

    blocks on a QUIC retrieve.

    blocks on a QUIC retrieve. First fetch after startup does a one-time

    synchronous wait so data is available for the first request.

  * Add revision guard: fetchFromCluster discards manifests with a

    revision lower than the cached revision, preventing data loss when

    a recovering node returns stale data.

  * Fix scrub data loss: scrub now validates retrieved data (size,

    AuthHeader marker) BEFORE calling remove(). Empty or corrupt

    retrieve results no longer trigger remove+store, which was deleting

    valid blocks from healthy nodes.

  * Add regression tests: PrefetchNeverBlocks, RevisionGuardNoDataLoss

 -- Jan Koester <jan.koester@tuxist.de>  Mon, 14 Apr 2026 22:00:00 +0200