Named dist_fun values are now limited to the curated list of measures returned by sc_list_distances(). Other measures can still be supplied as functions.
The curated set includes the new "cosine_distance" definition.
Added "wasserstein1" for probability profiles stored as ordered, equal-width raster layers.
"dtw2d" now rejects an odd number of layers instead of ignoring the final layer.
Invalid k values are now rejected before code is run.
Fixed R/C++ raster-index conversion for custom centers and metrics; metrics now preserve fractional center positions.
Initial center values are now read from the locally adjusted center cells.
use_meters() now warns explicitly that non-square cells use x resolution for both axes and therefore give approximate map-unit steps.
Fixed ID offsets and empty-result handling when combining chunked polygon outputs.
Balance metrics now return NA when either scaled distance component is zero.
Cluster assignments are now recomputed each iteration; valid cells outside local center windows are assigned to the nearest current center.
sc_slic_convergence() now skips connectivity cleaning and warns when clean = TRUE is requested.
Median summaries now average the two middle values for even-sized supercells.
supercells 1.9.5
Fixed integer overflow in vector indexing inside the C++ part of the code: C++ code now uses 64-bit arithmetic for array indices (std::size_t) instead of 32-bit integers when accessing raster data. This prevents silent data corruption or crashes when processing large rasters (>100M cells with multiple bands). The fix should have no performance impact. (Thanks to Lucas De Oto)
Chunking improvements: conservative memory estimates, options(supercells.chunk_mem_gb), chunk sizes align to step, deterministic ID offsets with file-backed merge
Verbose argument moved to the end in R and C++ APIs
Empty-cluster handling now consistent across cleanup
Removed future-based parallel chunking
supercells 1.6
Spatial distance uses precise (fractional) center positions; neighborhood search still uses rounded centers (minor output differences are expected vs 1.5)
More robust center handling (custom centers, empty centers) and safer cleanup for small regions
Local-minimum search improved
Tests standardized via tests/testthat/setup.R and updated expectations
supercells 1.5
Major C++ SLIC refactor (clearer data flow, diagnostics support) with supporting utility cleanup