Give Cp a noise variance without a nugget, seed from the scored parti…

NEWS.md

@@ -400,7 +400,7 @@
   `NA` with a `rejected_reason`) is not read as accepted: `cp` keeps that
   fit's nugget and warns, naming the reason, `reliability` is `NA` (it
   needs the range), and a fit that did not converge, or whose range is below
-  the shortest lag fitted, gives neither (the whole refused fit used to be
+  the shortest lag fitted, gives neither its nugget nor its range (the whole refused fit used to be
   used, and a correlation function whose range could be many times the
   extent pinned the reliability optimum to the first level, with no R
   warning).  A range below the shortest lag means the structure cannot be
@@ -422,8 +422,19 @@
   or a character string is refused by name; a `units` object used to be
   read as a number in the CRS units, so 1.5 km became a range of 1.5 m and
   a floor of 41 million cells.  A plain number is taken as the range alone:
-  it sets the floor, and `cp` and `reliability` are
-  `NA`.  Reliability's domain term is taken over the convex hull the
+  it sets the floor, and `reliability` is `NA`.  Wherever the variogram
+  gives no nugget (none was fitted, its fit did not converge, its range is
+  below the shortest lag, or `sac` is a range alone), `cp` takes Mallows'
+  own noise variance, the residual mean square RSS / (m - L) of the finest
+  level whose cells hold two scored rows each on average, and the profile
+  warns; `attr(x, "cp_noise")` records which variance was used, and the
+  print says so.  It counts the structure within those cells as noise too,
+  so on average it is no smaller than the nugget and errs towards fewer
+  cells.  `cp` used to be `NA` at every level in all four cases, so a
+  workflow that reads C_p by default (`build_tessellation(approx_n_cells =
+  <profile>)`, `summary()`) fell through to another criterion with a log
+  line at most.
+  Reliability's domain term is taken over the convex hull the
   area is measured on, not the bounding box: on a 3000 x 120 strip the
   reliability pick is 6 cells whether the strip lies axis-aligned or rotated
   by 45 degrees (it was 8 and 2), and 10 either way on a square (it was 10
@@ -862,8 +873,39 @@
   warns once for each model that did not converge.  Code that pins the
   exact column set of either table needs the new name added.
 
+* `select_resolution()` returns `$seeds`, the centres of the partition the
+  profile scored at the chosen level, and `get_voronoi_seeds(method =
+  "kmeans")` and `voronoi_seeds_kmeans()` return those centres when given
+  the selection or the profile, instead of running a k-means of their own.
+  A k-means partition is the Voronoi partition of its centres, so these
+  seeds rebuild the very cells the criteria judged.  The fresh k-means the
+  seeding functions ran (Hartigan-Wong, 10 starts, on whatever cloud they
+  were given) was a different partition: on 300 points it put 5--16 percent
+  of them in a different cell from the one they were scored in, and on
+  2000 points (profiled on a 1500-point subsample) 26--29 percent, so the
+  count was defended on cells nobody built.  The profile keeps every
+  level's centres in `attr(x, "centres")` (with `attr(x, "centres_crs")`),
+  the seeds come back in the CRS of the boundary or points they are for,
+  and `attr(seeds, "kmeans")` assigns `sample_points` to them, each point
+  to its nearest seed.  A count passed as a number (`n = sel$best`) still
+  runs a fresh k-means, as does a selection made before seeds were kept.
+  `voronoi_seeds_kmeans()` did not accept a selection or profile as `k`
+  before, and the advice in `build_tessellation()`'s occupancy warning now
+  points at the selection rather than the bare count.
+
 ## Bug fixes
 
+* **A clipped cell that also touched the boundary from outside was dropped,
+  and its points left with no cell.**  `st_intersection()` returns such a
+  cell as a GEOMETRYCOLLECTION of its area and a line or point (on an
+  L-shaped boundary, the cell over the inner corner), and the Voronoi, grid
+  and Delaunay builders kept only POLYGON and MULTIPOLYGON rows, so the
+  whole cell went, area and all.  On a 4 x 4 L with square cells of side 2,
+  ten of thirty points had no cell and the cells covered 8 of the
+  boundary's 10 square units.  Each collection is now reduced to its
+  polygonal part, one row per cell still; a piece with no area is dropped
+  as before.
+
 * **`coerce_to_points()` crashed R on an empty line feature.**  sf's
   `st_cast()` turns an empty MULTILINESTRING into one empty LINESTRING, not
   zero parts, and `st_line_sample()` on it segfaulted and took the session
@@ -1708,7 +1750,12 @@
   optimistic than the target.  `make_folds()` now warns when the floor
   leaves more than one point's worth of excess below `phi`, records
   `params$n_at_min_train`, and the documentation states the guarantee only
-  where neither `phi` nor `min_train` binds.
+  where neither `phi` nor `min_train` binds.  The warning names the
+  statistic that fires it, the largest excess of the realised
+  nearest-neighbour ECDF over the target at distances up to `phi`, and
+  gives the two medians only as context: they summarise all the distances,
+  and the realised median can sit above the target's while the short
+  distances are over-represented.
 
 * **`make_folds(auto_range = TRUE)` fell back to geometric blocks with only
   a log line.**  When no range was identified (an unremoved trend, a range

R/cross-validation.R

@@ -4701,17 +4701,25 @@ make_folds <- function(points_sf, k,
       rp <- rs[rs <= phi]
       max(findInterval(rp, rs) / length(rs) - G_target(rp))
     } else NA_real_
+    # The message states the statistic that fired it.  It used to lead with
+    # the two medians, which describe the whole distributions, not the short
+    # distances the test is about: the realised median could sit above the
+    # target's while the warning said the folds were "more optimistic".
     if (n_at_floor > 0L && is.finite(excess_phi) && excess_phi > 1 / n + 1e-9)
       .warn_and_log(paste0(
         "make_folds(nndm): min_train = %s stopped the distance matching in %d ",
         "of %d folds, which keep a training point closer than the prediction ",
-        "distances allow, so the realised distances remain more optimistic ",
-        "than the target (median %.4g against %.4g; largest ECDF excess ",
-        "%.3f). The prediction locations lie further from the samples than a ",
+        "distances allow. At distances up to phi (%.4g) the share of folds ",
+        "whose nearest training point is within a distance runs above the ",
+        "target's share by as much as %.3f (more than one fold's worth, 1/%d), ",
+        "so short distances are over-represented and the CV score is ",
+        "optimistic for the prediction locations. (The medians, %.4g realised ",
+        "and %.4g target, summarise all distances and need not differ the same ",
+        "way.) The prediction locations lie further from the samples than a ",
         "fold can be made to hold out: read the CV score as an upper bound on ",
         "performance there, or lower min_train."),
-        format(min_train), n_at_floor, n, stats::median(realised[fin]),
-        stats::median(g_target), excess_phi)
+        format(min_train), n_at_floor, n, phi, excess_phi, n,
+        stats::median(realised[fin]), stats::median(g_target))
 
     # Exclusion is not always needed.  When prediction locations sit no further
     # from the training data than training points sit from each other, plain