The marine-hull programme asks how far a small solar survey boat goes per day. Earlier steps showed a policy that slows down in rough water beats the best constant speed by 3 to 15%, but only with perfect foresight. Then a forecast of swept skill recovered under half of that.
Milestone 6e uses a forecast with no free parameter: today's call is yesterday's sea state. With rough spells averaging seven days, that is right 85% of the time, and it recovers 68 to 80% of the gap in all ten cells (five sites, two battery sizes). Its errors fall on spell boundaries and cost one day each, which is why it beats the independent-error forecast of the previous step at a similar hit rate.
One result is an artifact, and the log says so. The one-day-spell cells at p = 0.5 come from a Markov chain that alternates deterministically, so yesterday predicts today perfectly inverted. Those cells show persistence equal to the oracle, and they should not be quoted. Sea state itself is still a swept two-state model, not a wave climatology.
Log and data: research/marine-hull/LOG.md, experiments/results_solar_seastate_persist.txt.