windfarmGA

• Legacy crossover(), mutation() and trimton() still accept 0/1 chromosomes. New exports: set_crossover(), swap_mutation().
• plot_fitness_evolution() is gone (it only called plot_parkfitness()).

New

• Results have class windfarmGA (print / plot). ga_options() lists or sets options(windfarmGA.*).
• explore_result(): Shiny map of a generation, wind rose, downwind wake cones, optional terrain layers + turbine popups. Click a new fitness max to jump the slider.
• Fitness is \(E \times (\eta/100)^w\) (windfarmGA.fitness_efficiency_weight, default 1). Log hub-height profile (power law via windfarmGA.wind_profile = "power"). Optional manufacturer curve (ga_options(power_curve = …), read_power_curve(), plot_power_curve()). Cp default 0.45; optional cut-in / rated / cut-out.
• Elitism archives the best layout (no fitness × 10). Elite children, neighbour local search, fitness cache, immigrants, spatial crossover. Early stop only after stall_generations idle gens (no new layout, cell, or best).
• Terrain: pass a DEM (terrain = dem) to skip the download. Per-cell height / \(z_0\) / \(k\) / air density are computed once and stored as terrainModel for plot_result and random_search.
• random_search() jitters turbines inside their cells (runs). Terrain is reused from result; pass weibull_src again (not stored). plot_random_search() draws both panels on the same grid.
• Wake search and circle overlap run in C++. Turbines exactly upwind (alpha = 0) are kept (old triangle test dropped them).
• Wind helpers: wind_from_uv(), wind_from_series(). GitHub-only extras in experimental/ (ERA5/GWA, noise, rayshader, mapgl) are not in the CRAN tarball. plot_mapgl_from_result() reuses terrain_tiles/ / terrain.mbtiles, draws wake cones colored by AbschGesamt (0% green → max red) and the rayshader OBJ (wind_turbine_v1.obj). Three.js is r149 UMD from localhost. TileJSON Terrarium encoding; maxzoom is the highest tile folder. Basemaps: satellite, topo, Carto. Port 8000 leftover servers are freed (rebuild = TRUE remakes the tiles).
• Plots wait for Enter between pages when interactive. plot_generation() shows every layout in a generation. plot_viewshed() projects lon/lat DEMs first.
• genetic_algorithm(on_generation = ...) calls function(generation, iteration, energy, efficiency, fitness) after each generation so a Shiny app can show progress without forking the GA. wake_cones() is the public name for the downwind search-cone polygons (plot_leaflet, MapGL). Internal geometry is wake_cone_polys().

Fixes

• Weibull speed no longer becomes all-NA after masking (fitness crash).
• Power-curve energy is the park sum, not the first turbine.
• clc_legend.csv is in the tarball again (.Rbuildignore had dropped every CSV).
• crossover parent fitness is (a + b) / 2. isSpatial uses the given CRS, not hardcoded 3035.
• plot_windrose() works with ggplot2 4.x (S7). If ggplot2 4 meets an old systemfonts, update systemfonts.
• Parallel clusters always stop via on.exit.
• plot_result(terrain = dem) no longer errors on a SpatRaster. Weibull file paths get a CRS before crop. PROJ < 6 branches are gone.
• with_mocked_bindings() in test_coverage_branches.R wraps the ggplot2-missing checks in code = { … } (testthat requires named bindings). random_search() test uses runs, not the removed n.
• resultrect is the first 50 generations of the old 200-iteration run, saved with xz (~90 KB). A new 50-gen GA is larger because allCoords now stores a bigger population.
• plot_cell_heatmap() / ga_result_grid() no longer assume a missing CRS is lon/lat. Meter coordinates get the result projection (same as plot_leaflet()). The 4326 round-trip broke the grid on macOS.
• Roxygen: unused @inheritParams on trimton() and turbine_influences() removed (params were already documented).
• Rd help is ASCII-only (≈, ×, em-dash broke the CRAN PDF manual). README no longer links to the archived check-results page (404).
• Parallel: n_cluster is capped at 2 under _R_CHECK_LIMIT_CORES_ (R errors on makeCluster(detectCores() - 1)). Terrain tests no longer need a local g100_06.tif.
• random_search_single(turbine = …) skips the interactive prompt. User-input tests no longer write to a fake stdin.
• stopCluster() is wrapped in try() so a dead worker does not fail the run. Live parallel GA tests skip under covr. Time sequential vs PSOCK locally with source("experimental/bench_parallel.R"); bench_parallel(). Windows starts PSOCK workers sequentially; fitness() loads windfarmGA/terra/sf on the workers. The bench catches dropped connections (often at 6–8 workers on this small site). README notes ~4 workers as the useful default on the demo site.
• calculate_energy() uses its public arguments (layout, area, wind, …) instead of internal aliases. Unused maxt0 in the GA first-generation fuzzy block is gone.
• windata_format() column matching used a broken sapply(grepl, …) (--run-donttest example). Frames that already have ws/wd skip the matcher (so probab no longer triggers it).