Running tally — median |rate| vs. interval length
What you just did has a name
Rounds 1–3 showed the same LTEE population, the same 50,000 generations of fitness data. You only changed the length of the sliding window over which you computed a rate. The median rate dropped as the window grew. Nothing changed about the biology.
This is Gingerich's 1983 result: rates of evolution decline with the interval over which they are measured. Part of the decline is a real biological fact (evolution proceeds in bursts separated by stasis, so longer windows average burst + stasis together). Part of it is a statistical artifact (measurement noise and short-term reversals average out over longer intervals). Without a null model you cannot separate the two.
The scale of the decline is enormous. LTEE at 500 generations gives rates orders of magnitude larger than Grant finches at 1 year, which are in turn orders of magnitude larger than Hyopsodus over 500,000 years. Published rate compilations (Gingerich 2019, Hendry & Kinnison 1999) show the same pattern across hundreds of studies. A single "rate of evolution" has no meaning without an interval. Always report the interval.