Staleness Handling

Removing staleness decay reduced final accuracy and increased time-to-target in ablation experiments. Reflective Async achieved the best evolution efficiency in IFBench ablations. Full Async maximizes throughput by allowing all items to continue but can admit outdated updates into the pool. The convergence analysis treats stale updates as introducing a bias term that grows with the square of the maximum staleness. FedQueue admits updates by a round cutoff and buffers late updates for a later cutoff instead of discarding them. Queue items carry an artifact-pool version so FlashEvolve can compare item age against the current pool version. Reflective Async inspects stale items and intervening pool updates to accept, patch, or discard them. Staleness decay reduces the influence of delayed updates without discarding them. FedQueue applies harmonic staleness decay by default when aggregating updates. Guarded Async discards items whose version gap exceeds a threshold, trading reduced staleness for wasted generated tokens. Artifact staleness arises because queued intermediate results may have been produced from older artifact-pool versions.