downlit vs Dragonfly

The last two months are almost entirely correctness work: a pub/sub ordering bug that silently discarded queued messages on unsubscribe, a multi-key PFCOUNT that read sixteen wrong registers, an uninitialized pointer in the sparse HyperLogLog promote path, and now cluster FLUSHSLOTS that was not journaled inside its own flush transaction. The one structural move is the resync of the forked hyperloglog.c against valkey upstream, which cut divergence from 2,139 lines to 767 and pulled in a CVE fix plus AVX2 and NEON merge kernels. New user-facing options are rare; the list compression tuning flags are the exception.

Two threads run in parallel. One reduces the cost of carrying a Redis fork by re-syncing individual files against valkey and leaving provenance notes so the next sync is cheap. The other hardens the V2 IO loop while the V1 path still takes the fixes — several commits explicitly defer the V2 half of a fix to a later change, so the two loops are not at parity. Cluster correctness around slot migration and journaling is the newest front.

The deferred V2 IO-loop fixes are the obvious next commits, and more forked Redis files are likely to get the same upstream resync treatment now that the pattern and its accounting are established.