A sail drone that replans hourly from a poor wind forecast tacks 150 to 300 times a day and loses to a rule that just follows the wind shifts. Two earlier steps tried a switching margin and found it did not help. This step tried the other obvious fix: forbid a voluntary tack within D minutes of the last one.
Using real NOAA CO-OPS wind records (January and July 2024, 48 windows each), the rule did cut tacks, by 55% in January and 39% in July at zero forecast skill. It did not buy progress. With no forecast skill every dwell value did worse than the header-follower (by 1.8 to 3.9 points of the bound). With good skill (120 minutes) even a 10 minute dwell cost about 1.7 points, and 40 minutes cost 4 to 5.
So the tack count is not what costs a replanner progress; the forecast is. Margin and dwell damping are now both closed off, and the header-follower is the better rule below roughly 30 minutes of skill.
The forecast error is a synthetic model, and a tack costs only a 30 s penalty here. Numbers: research/marine-hull/experiments/results_sail_dwell.txt.