RubyGems - carray - Versions diffs - 3.0.2 → 3.0.4 - Mend

carray 3.0.2 → 3.0.4

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.

Files changed (247) hide show

checksums.yaml +4 -4
data/CHANGELOG.md +568 -0
data/README.md +6 -6
data/carray.gemspec +1 -1
data/ext/ca_axis_descriptor.h +27 -24
data/ext/ca_axis_dispatch.c +220 -0
data/ext/ca_axis_group.c +201 -123
data/ext/ca_binop_dispatch.c +46 -52
data/ext/ca_binop_dispatch.h +1 -1
data/ext/ca_categorical_iterator.c +249 -183
data/ext/ca_compare.c +7 -12
data/ext/ca_for_buffer.h +21 -68
data/ext/ca_group_iter.c +38 -23
data/ext/ca_kernel_iterator.c +1765 -2076
data/ext/ca_kernel_iterator.h +584 -566
data/ext/ca_monop_dispatch.c +86 -63
data/ext/ca_monop_dispatch.h +5 -0
data/ext/ca_obj_array.c +223 -56
data/ext/ca_obj_bincmp.c +172 -138
data/ext/ca_obj_binop.c +116 -230
data/ext/ca_obj_bitarray.c +40 -26
data/ext/ca_obj_bitfield.c +100 -68
data/ext/ca_obj_block.c +19 -9
data/ext/ca_obj_byte_swap.c +45 -46
data/ext/ca_obj_const_string.c +269 -44
data/ext/ca_obj_face.c +612 -89
data/ext/ca_obj_face.h +111 -47
data/ext/ca_obj_fake.c +36 -79
data/ext/ca_obj_farray.c +7 -3
data/ext/ca_obj_field.c +33 -8
data/ext/ca_obj_fixlen_string.c +23 -24
data/ext/ca_obj_grid.c +541 -210
data/ext/ca_obj_meld.c +331 -111
data/ext/ca_obj_moncmp.c +80 -65
data/ext/ca_obj_monop.c +177 -149
data/ext/ca_obj_object.c +664 -248
data/ext/ca_obj_record.c +63 -23
data/ext/ca_obj_reduce.c +18 -16
data/ext/ca_obj_refer.c +140 -58
data/ext/ca_obj_remap.c +85 -24
data/ext/ca_obj_repeat.c +24 -9
data/ext/ca_obj_roll.c +121 -16
data/ext/ca_obj_select.c +61 -58
data/ext/ca_obj_select_axis.c +73 -107
data/ext/ca_obj_shift.c +32 -102
data/ext/ca_obj_stack.c +372 -122
data/ext/ca_obj_stride.c +348 -140
data/ext/ca_obj_string.c +88 -22
data/ext/ca_obj_tile.c +38 -21
data/ext/ca_obj_time.c +45 -31
data/ext/ca_obj_timedelta.c +43 -29
data/ext/ca_obj_transpose.c +17 -6
data/ext/ca_obj_triop.c +125 -141
data/ext/ca_obj_window.c +330 -44
data/ext/ca_op_ipower.c +24 -11
data/ext/ca_op_powi.h +57 -30
data/ext/ca_pow5_table.h +656 -0
data/ext/ca_sort_kernels.h +70 -0
data/ext/ca_sweep_engine.c +104 -81
data/ext/ca_sweep_engine.h +7 -20
data/ext/ca_triop_dispatch.c +4 -1
data/ext/ca_triop_dispatch.h +5 -2
data/ext/caframe_csv_reader.c +542 -0
data/ext/carray.h +241 -92
data/ext/carray_access.c +608 -276
data/ext/carray_address_basis.c +494 -97
data/ext/carray_attribute.c +129 -4
data/ext/carray_bincount.c +39 -7
data/ext/carray_broadcast.c +56 -19
data/ext/carray_build_flags.h +1 -1
data/ext/carray_call_cfunc.c +108 -28
data/ext/carray_cast.c +1576 -223
data/ext/carray_cast_func.rb +77 -24
data/ext/carray_class.c +4 -3
data/ext/carray_conversion.c +208 -75
data/ext/carray_copy.c +82 -24
data/ext/carray_core.c +646 -162
data/ext/carray_count.c +171 -88
data/ext/carray_element.c +94 -60
data/ext/carray_factorize.c +670 -324
data/ext/carray_generate.c +260 -83
data/ext/carray_histogram.c +209 -106
data/ext/carray_hold.c +14 -16
data/ext/carray_index_classifier.c +422 -243
data/ext/carray_internal.h +692 -3
data/ext/carray_kernels_bincmp.c +7 -0
data/ext/carray_kernels_binop.c +487 -120
data/ext/carray_kernels_init.c +653 -50
data/ext/carray_kernels_moncmp.c +15 -8
data/ext/carray_kernels_monop.c +99 -43
data/ext/carray_kernels_reduce_aggregate.c +3925 -6112
data/ext/carray_kernels_reduce_boolean.c +112 -36
data/ext/carray_kernels_reduce_cumulative.c +2127 -6903
data/ext/carray_kernels_reduce_extreme.c +3032 -2892
data/ext/carray_kernels_reduce_variance.c +439 -500
data/ext/carray_kernels_scan.c +1097 -488
data/ext/carray_kernels_search.c +3787 -2176
data/ext/carray_kernels_sort.c +1200 -920
data/ext/carray_kernels_triop.c +479 -7
data/ext/carray_key_windows.c +216 -0
data/ext/carray_lazy.c +967 -153
data/ext/carray_loop.c +81 -65
data/ext/carray_mask.c +910 -324
data/ext/carray_math_kernel.h +1 -0
data/ext/carray_median_percentile.c +299 -100
data/ext/carray_memory_view.c +872 -758
data/ext/carray_operator.c +1171 -364
data/ext/carray_order.c +505 -133
data/ext/carray_partition.c +112 -65
data/ext/carray_random.c +327 -239
data/ext/carray_round.c +273 -0
data/ext/carray_scatter.c +346 -78
data/ext/carray_segment.c +257 -0
data/ext/carray_slab.c +249 -86
data/ext/carray_sort.c +224 -227
data/ext/carray_sort_kernel.c +8 -1
data/ext/carray_struct.c +118 -39
data/ext/carray_test.c +193 -69
data/ext/carray_time_text.c +1182 -0
data/ext/carray_undef.c +4 -0
data/ext/carray_utils.c +579 -113
data/ext/extconf.rb +27 -2
data/ext/mk_call_cfunc.rb +19 -4
data/ext/mk_pow5_table.rb +35 -0
data/ext/mkkernel.rb +1202 -1014
data/ext/ruby_carray.c +17 -5
data/ext/version.h +4 -4
data/ext/version.rb +4 -2
data/lib/carray/attribute.rb +160 -17
data/lib/carray/autoload_carray.rb +23 -27
data/lib/carray/autoload_method_extension.rb +20 -5
data/lib/carray/axis_group.rb +360 -63
data/lib/carray/basics.rb +162 -28
data/lib/carray/bincount_nd.rb +18 -2
data/lib/carray/block_iterator.rb +231 -58
data/lib/carray/boolean_reduce.rb +34 -42
data/lib/carray/categorical.rb +242 -52
data/lib/carray/categorical_iterator.rb +437 -827
data/lib/carray/complex.rb +41 -34
data/lib/carray/conditional.rb +150 -14
data/lib/carray/const_string.rb +59 -4
data/lib/carray/construct.rb +15 -11
data/lib/carray/data_type_extension.rb +1 -0
data/lib/carray/distance_limit.rb +77 -0
data/lib/carray/fixlen_string.rb +7 -1
data/lib/carray/frame/column_verbs.rb +70 -0
data/lib/carray/frame/concat.rb +157 -6
data/lib/carray/frame/convert.rb +82 -7
data/lib/carray/frame/csv_parser.rb +688 -87
data/lib/carray/frame/describe.rb +80 -0
data/lib/carray/frame/frame.rb +369 -45
data/lib/carray/frame/group.rb +631 -90
data/lib/carray/frame/io.rb +506 -79
data/lib/carray/frame/join.rb +411 -62
data/lib/carray/frame/records.rb +116 -19
data/lib/carray/frame/reshape.rb +455 -0
data/lib/carray/frame/sort.rb +30 -18
data/lib/carray/frame/verbs.rb +868 -61
data/lib/carray/frame/windows.rb +117 -0
data/lib/carray/frame.rb +4 -0
data/lib/carray/fuse_source.rb +118 -2
data/lib/carray/fusion.rb +257 -42
data/lib/carray/histogram.rb +6 -1
data/lib/carray/inspect.rb +54 -13
data/lib/carray/iterator.rb +99 -9
data/lib/carray/key_windows.rb +620 -0
data/lib/carray/lazy.rb +216 -193
data/lib/carray/mask_gap_fill.rb +112 -92
data/lib/carray/math.rb +32 -19
data/lib/carray/meld_reduce.rb +52 -21
data/lib/carray/methods/align_addr.rb +8 -4
data/lib/carray/methods/bin.rb +26 -18
data/lib/carray/methods/bincount.rb +18 -2
data/lib/carray/methods/bit_string.rb +15 -4
data/lib/carray/methods/choose.rb +16 -3
data/lib/carray/methods/composition.rb +18 -10
data/lib/carray/methods/factorize.rb +2 -2
data/lib/carray/methods/gather_nd.rb +25 -5
data/lib/carray/methods/insert_block.rb +9 -14
data/lib/carray/methods/is_between.rb +59 -0
data/lib/carray/methods/is_in.rb +140 -10
data/lib/carray/methods/join.rb +2 -7
data/lib/carray/methods/lag.rb +49 -0
data/lib/carray/methods/locate_addr.rb +182 -45
data/lib/carray/methods/mode.rb +18 -1
data/lib/carray/methods/nunique.rb +11 -6
data/lib/carray/methods/repeat.rb +1 -20
data/lib/carray/methods/resize.rb +21 -23
data/lib/carray/methods/snap.rb +34 -12
data/lib/carray/methods/unique.rb +11 -2
data/lib/carray/methods/value_counts.rb +4 -11
data/lib/carray/record_order.rb +208 -0
data/lib/carray/rng.rb +3 -0
data/lib/carray/runtime.rb +65 -17
data/lib/carray/segment_iterator.rb +1221 -0
data/lib/carray/serialize.rb +214 -30
data/lib/carray/slab_iterator.rb +101 -27
data/lib/carray/stack.rb +10 -4
data/lib/carray/string.rb +105 -5
data/lib/carray/string_operation_extension.rb +12 -4
data/lib/carray/struct.rb +137 -58
data/lib/carray/struct_builder.rb +124 -17
data/lib/carray/time.rb +1222 -186
data/lib/carray/window_iterator.rb +242 -81
data/lib/carray.rb +9 -2
data/yard-stubs/ca_obj_array.rb +4 -4
data/yard-stubs/ca_obj_bitarray.rb +7 -0
data/yard-stubs/ca_obj_bitfield.rb +23 -12
data/yard-stubs/ca_obj_fake.rb +1 -1
data/yard-stubs/ca_obj_field.rb +5 -1
data/yard-stubs/ca_obj_grid.rb +21 -10
data/yard-stubs/ca_obj_refer.rb +5 -4
data/yard-stubs/ca_obj_roll.rb +2 -3
data/yard-stubs/ca_obj_shift.rb +5 -1
data/yard-stubs/ca_obj_stride.rb +10 -7
data/yard-stubs/ca_obj_tile.rb +1 -1
data/yard-stubs/ca_obj_transpose.rb +1 -1
data/yard-stubs/ca_obj_window.rb +9 -5
data/yard-stubs/carray_access.rb +11 -5
data/yard-stubs/carray_attribute.rb +0 -5
data/yard-stubs/carray_broadcast.rb +1 -1
data/yard-stubs/carray_cast.rb +88 -43
data/yard-stubs/carray_conversion.rb +15 -2
data/yard-stubs/carray_copy.rb +12 -5
data/yard-stubs/carray_core.rb +3 -3
data/yard-stubs/carray_count.rb +9 -0
data/yard-stubs/carray_element.rb +10 -4
data/yard-stubs/carray_generate.rb +21 -1
data/yard-stubs/carray_lazy.rb +5 -1
data/yard-stubs/carray_loop.rb +4 -1
data/yard-stubs/carray_mask.rb +60 -12
data/yard-stubs/carray_math.rb +52 -14
data/yard-stubs/carray_median_percentile.rb +15 -8
data/yard-stubs/carray_memory_view.rb +40 -23
data/yard-stubs/carray_order.rb +61 -29
data/yard-stubs/carray_random.rb +13 -4
data/yard-stubs/carray_scatter.rb +21 -9
data/yard-stubs/carray_segment.rb +72 -0
data/yard-stubs/carray_slab.rb +10 -2
data/yard-stubs/carray_sort.rb +49 -8
data/yard-stubs/carray_test.rb +15 -3
data/yard-stubs/carray_utils.rb +0 -26
metadata +19 -5
data/ext/ca_for_each_element.h +0 -239
data/ext/carray_kernels_map.c +0 -3467
data/lib/carray/arrow_tensor.rb +0 -401
data/lib/carray/table.rb +0 -108

data/ext/carray_kernels_reduce_variance.c CHANGED Viewed

@@ -24,6 +24,13 @@
    probed by operation-table function pointer (obj_type is runtime-
    assigned, so no CA_OBJ_STACK compile-time constant exists). */
 extern ca_operation_function_t ca_stack_func;
+/* The square root of a variance: an ε-negative from rounding becomes 0,
+   a NaN stays NaN (fmax(NaN, 0.0) is 0.0). */
+#ifndef CARRAY_SQRT_CLAMP_DEFINED
+#define CARRAY_SQRT_CLAMP_DEFINED
+static inline double ca_sqrt_clamp (double v) { return sqrt(v < 0.0 ? 0.0 : v); }
+#endif
 /* BOOL2VAL: bool -> Ruby (Qtrue/Qfalse).  Used as ruby_scalar wrapper
    for the all/any flat-reduction Ruby surface so `a.all` / `a.any`
    return real true/false, not Integer 0/1.  (BOOL2OBJ already exists
@@ -52,7 +59,7 @@ variancep_ki_native_i8 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes,
                                  slab_axes, naxes, 0);
   if ( rc != CA_ITER_OK ) {
     rb_raise(rb_eRuntimeError,
-             "variancep_ki: kernel_iterator init failed rc=%d", rc);
+             "%s: kernel_iterator init failed rc=%d", ca_calling_method_name(), rc);
   }
   char       *p;
   boolean8_t *m;
@@ -64,7 +71,7 @@ variancep_ki_native_i8 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes,
     ca_size_t masked_cnt = 0;
     CA_SLAB_REDUCE_SUM8_EX(int8_t, double, st, p, m, sum, 0.0, __variancep_i8_P1EXPR, masked_cnt);
     ca_size_t n_valid = st.slab_elements - masked_cnt;
-    int __trigger = (min_count < 0 ? masked_cnt == st.slab_elements : n_valid < min_count);
+    int __trigger = ca_too_few_present(n_valid, min_count, 1);
     if ( __trigger ) {
       if ( ! op_mask ) {
         ca_create_mask(co);
@@ -72,11 +79,6 @@ variancep_ki_native_i8 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes,
       }
       op_mask[out_i] = 1;
       op[out_i++] = (double) 0.0;
-    } else if ( n_valid < 1 ) {
-      /* Legacy degenerate case (e.g. sample variance with 1 unmasked
-         cell): return zero rather than UNDEF, matching pre-migration
-         behaviour of `cnt > 1 ? formula : 0`. */
-      op[out_i++] = (double) 0.0;
     } else {
       double mean = sum / (double) n_valid;
       /* Pass 2: M2 = Σ (x - mean)² */
@@ -114,7 +116,7 @@ variancep_ki_native_u8 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes,
                                  slab_axes, naxes, 0);
   if ( rc != CA_ITER_OK ) {
     rb_raise(rb_eRuntimeError,
-             "variancep_ki: kernel_iterator init failed rc=%d", rc);
+             "%s: kernel_iterator init failed rc=%d", ca_calling_method_name(), rc);
   }
   char       *p;
   boolean8_t *m;
@@ -126,7 +128,7 @@ variancep_ki_native_u8 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes,
     ca_size_t masked_cnt = 0;
     CA_SLAB_REDUCE_SUM8_EX(uint8_t, double, st, p, m, sum, 0.0, __variancep_u8_P1EXPR, masked_cnt);
     ca_size_t n_valid = st.slab_elements - masked_cnt;
-    int __trigger = (min_count < 0 ? masked_cnt == st.slab_elements : n_valid < min_count);
+    int __trigger = ca_too_few_present(n_valid, min_count, 1);
     if ( __trigger ) {
       if ( ! op_mask ) {
         ca_create_mask(co);
@@ -134,11 +136,6 @@ variancep_ki_native_u8 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes,
       }
       op_mask[out_i] = 1;
       op[out_i++] = (double) 0.0;
-    } else if ( n_valid < 1 ) {
-      /* Legacy degenerate case (e.g. sample variance with 1 unmasked
-         cell): return zero rather than UNDEF, matching pre-migration
-         behaviour of `cnt > 1 ? formula : 0`. */
-      op[out_i++] = (double) 0.0;
     } else {
       double mean = sum / (double) n_valid;
       /* Pass 2: M2 = Σ (x - mean)² */
@@ -176,7 +173,7 @@ variancep_ki_native_i16 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes
                                  slab_axes, naxes, 0);
   if ( rc != CA_ITER_OK ) {
     rb_raise(rb_eRuntimeError,
-             "variancep_ki: kernel_iterator init failed rc=%d", rc);
+             "%s: kernel_iterator init failed rc=%d", ca_calling_method_name(), rc);
   }
   char       *p;
   boolean8_t *m;
@@ -188,7 +185,7 @@ variancep_ki_native_i16 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes
     ca_size_t masked_cnt = 0;
     CA_SLAB_REDUCE_SUM8_EX(int16_t, double, st, p, m, sum, 0.0, __variancep_i16_P1EXPR, masked_cnt);
     ca_size_t n_valid = st.slab_elements - masked_cnt;
-    int __trigger = (min_count < 0 ? masked_cnt == st.slab_elements : n_valid < min_count);
+    int __trigger = ca_too_few_present(n_valid, min_count, 1);
     if ( __trigger ) {
       if ( ! op_mask ) {
         ca_create_mask(co);
@@ -196,11 +193,6 @@ variancep_ki_native_i16 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes
       }
       op_mask[out_i] = 1;
       op[out_i++] = (double) 0.0;
-    } else if ( n_valid < 1 ) {
-      /* Legacy degenerate case (e.g. sample variance with 1 unmasked
-         cell): return zero rather than UNDEF, matching pre-migration
-         behaviour of `cnt > 1 ? formula : 0`. */
-      op[out_i++] = (double) 0.0;
     } else {
       double mean = sum / (double) n_valid;
       /* Pass 2: M2 = Σ (x - mean)² */
@@ -238,7 +230,7 @@ variancep_ki_native_u16 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes
                                  slab_axes, naxes, 0);
   if ( rc != CA_ITER_OK ) {
     rb_raise(rb_eRuntimeError,
-             "variancep_ki: kernel_iterator init failed rc=%d", rc);
+             "%s: kernel_iterator init failed rc=%d", ca_calling_method_name(), rc);
   }
   char       *p;
   boolean8_t *m;
@@ -250,7 +242,7 @@ variancep_ki_native_u16 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes
     ca_size_t masked_cnt = 0;
     CA_SLAB_REDUCE_SUM8_EX(uint16_t, double, st, p, m, sum, 0.0, __variancep_u16_P1EXPR, masked_cnt);
     ca_size_t n_valid = st.slab_elements - masked_cnt;
-    int __trigger = (min_count < 0 ? masked_cnt == st.slab_elements : n_valid < min_count);
+    int __trigger = ca_too_few_present(n_valid, min_count, 1);
     if ( __trigger ) {
       if ( ! op_mask ) {
         ca_create_mask(co);
@@ -258,11 +250,6 @@ variancep_ki_native_u16 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes
       }
       op_mask[out_i] = 1;
       op[out_i++] = (double) 0.0;
-    } else if ( n_valid < 1 ) {
-      /* Legacy degenerate case (e.g. sample variance with 1 unmasked
-         cell): return zero rather than UNDEF, matching pre-migration
-         behaviour of `cnt > 1 ? formula : 0`. */
-      op[out_i++] = (double) 0.0;
     } else {
       double mean = sum / (double) n_valid;
       /* Pass 2: M2 = Σ (x - mean)² */
@@ -300,7 +287,7 @@ variancep_ki_native_i32 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes
                                  slab_axes, naxes, 0);
   if ( rc != CA_ITER_OK ) {
     rb_raise(rb_eRuntimeError,
-             "variancep_ki: kernel_iterator init failed rc=%d", rc);
+             "%s: kernel_iterator init failed rc=%d", ca_calling_method_name(), rc);
   }
   char       *p;
   boolean8_t *m;
@@ -312,7 +299,7 @@ variancep_ki_native_i32 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes
     ca_size_t masked_cnt = 0;
     CA_SLAB_REDUCE_SUM8_EX(int32_t, double, st, p, m, sum, 0.0, __variancep_i32_P1EXPR, masked_cnt);
     ca_size_t n_valid = st.slab_elements - masked_cnt;
-    int __trigger = (min_count < 0 ? masked_cnt == st.slab_elements : n_valid < min_count);
+    int __trigger = ca_too_few_present(n_valid, min_count, 1);
     if ( __trigger ) {
       if ( ! op_mask ) {
         ca_create_mask(co);
@@ -320,11 +307,6 @@ variancep_ki_native_i32 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes
       }
       op_mask[out_i] = 1;
       op[out_i++] = (double) 0.0;
-    } else if ( n_valid < 1 ) {
-      /* Legacy degenerate case (e.g. sample variance with 1 unmasked
-         cell): return zero rather than UNDEF, matching pre-migration
-         behaviour of `cnt > 1 ? formula : 0`. */
-      op[out_i++] = (double) 0.0;
     } else {
       double mean = sum / (double) n_valid;
       /* Pass 2: M2 = Σ (x - mean)² */
@@ -362,7 +344,7 @@ variancep_ki_native_u32 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes
                                  slab_axes, naxes, 0);
   if ( rc != CA_ITER_OK ) {
     rb_raise(rb_eRuntimeError,
-             "variancep_ki: kernel_iterator init failed rc=%d", rc);
+             "%s: kernel_iterator init failed rc=%d", ca_calling_method_name(), rc);
   }
   char       *p;
   boolean8_t *m;
@@ -374,7 +356,7 @@ variancep_ki_native_u32 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes
     ca_size_t masked_cnt = 0;
     CA_SLAB_REDUCE_SUM8_EX(uint32_t, double, st, p, m, sum, 0.0, __variancep_u32_P1EXPR, masked_cnt);
     ca_size_t n_valid = st.slab_elements - masked_cnt;
-    int __trigger = (min_count < 0 ? masked_cnt == st.slab_elements : n_valid < min_count);
+    int __trigger = ca_too_few_present(n_valid, min_count, 1);
     if ( __trigger ) {
       if ( ! op_mask ) {
         ca_create_mask(co);
@@ -382,11 +364,6 @@ variancep_ki_native_u32 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes
       }
       op_mask[out_i] = 1;
       op[out_i++] = (double) 0.0;
-    } else if ( n_valid < 1 ) {
-      /* Legacy degenerate case (e.g. sample variance with 1 unmasked
-         cell): return zero rather than UNDEF, matching pre-migration
-         behaviour of `cnt > 1 ? formula : 0`. */
-      op[out_i++] = (double) 0.0;
     } else {
       double mean = sum / (double) n_valid;
       /* Pass 2: M2 = Σ (x - mean)² */
@@ -424,7 +401,7 @@ variancep_ki_native_i64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes
                                  slab_axes, naxes, 0);
   if ( rc != CA_ITER_OK ) {
     rb_raise(rb_eRuntimeError,
-             "variancep_ki: kernel_iterator init failed rc=%d", rc);
+             "%s: kernel_iterator init failed rc=%d", ca_calling_method_name(), rc);
   }
   char       *p;
   boolean8_t *m;
@@ -436,7 +413,7 @@ variancep_ki_native_i64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes
     ca_size_t masked_cnt = 0;
     CA_SLAB_REDUCE_SUM8_EX(int64_t, double, st, p, m, sum, 0.0, __variancep_i64_P1EXPR, masked_cnt);
     ca_size_t n_valid = st.slab_elements - masked_cnt;
-    int __trigger = (min_count < 0 ? masked_cnt == st.slab_elements : n_valid < min_count);
+    int __trigger = ca_too_few_present(n_valid, min_count, 1);
     if ( __trigger ) {
       if ( ! op_mask ) {
         ca_create_mask(co);
@@ -444,11 +421,6 @@ variancep_ki_native_i64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes
       }
       op_mask[out_i] = 1;
       op[out_i++] = (double) 0.0;
-    } else if ( n_valid < 1 ) {
-      /* Legacy degenerate case (e.g. sample variance with 1 unmasked
-         cell): return zero rather than UNDEF, matching pre-migration
-         behaviour of `cnt > 1 ? formula : 0`. */
-      op[out_i++] = (double) 0.0;
     } else {
       double mean = sum / (double) n_valid;
       /* Pass 2: M2 = Σ (x - mean)² */
@@ -486,7 +458,7 @@ variancep_ki_native_u64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes
                                  slab_axes, naxes, 0);
   if ( rc != CA_ITER_OK ) {
     rb_raise(rb_eRuntimeError,
-             "variancep_ki: kernel_iterator init failed rc=%d", rc);
+             "%s: kernel_iterator init failed rc=%d", ca_calling_method_name(), rc);
   }
   char       *p;
   boolean8_t *m;
@@ -498,7 +470,7 @@ variancep_ki_native_u64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes
     ca_size_t masked_cnt = 0;
     CA_SLAB_REDUCE_SUM8_EX(uint64_t, double, st, p, m, sum, 0.0, __variancep_u64_P1EXPR, masked_cnt);
     ca_size_t n_valid = st.slab_elements - masked_cnt;
-    int __trigger = (min_count < 0 ? masked_cnt == st.slab_elements : n_valid < min_count);
+    int __trigger = ca_too_few_present(n_valid, min_count, 1);
     if ( __trigger ) {
       if ( ! op_mask ) {
         ca_create_mask(co);
@@ -506,11 +478,6 @@ variancep_ki_native_u64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes
       }
       op_mask[out_i] = 1;
       op[out_i++] = (double) 0.0;
-    } else if ( n_valid < 1 ) {
-      /* Legacy degenerate case (e.g. sample variance with 1 unmasked
-         cell): return zero rather than UNDEF, matching pre-migration
-         behaviour of `cnt > 1 ? formula : 0`. */
-      op[out_i++] = (double) 0.0;
     } else {
       double mean = sum / (double) n_valid;
       /* Pass 2: M2 = Σ (x - mean)² */
@@ -548,7 +515,7 @@ variancep_ki_native_f32 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes
                                  slab_axes, naxes, 0);
   if ( rc != CA_ITER_OK ) {
     rb_raise(rb_eRuntimeError,
-             "variancep_ki: kernel_iterator init failed rc=%d", rc);
+             "%s: kernel_iterator init failed rc=%d", ca_calling_method_name(), rc);
   }
   char       *p;
   boolean8_t *m;
@@ -560,7 +527,7 @@ variancep_ki_native_f32 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes
     ca_size_t masked_cnt = 0;
     CA_SLAB_REDUCE_SUM8_EX(float, double, st, p, m, sum, 0.0, __variancep_f32_P1EXPR, masked_cnt);
     ca_size_t n_valid = st.slab_elements - masked_cnt;
-    int __trigger = (min_count < 0 ? masked_cnt == st.slab_elements : n_valid < min_count);
+    int __trigger = ca_too_few_present(n_valid, min_count, 1);
     if ( __trigger ) {
       if ( ! op_mask ) {
         ca_create_mask(co);
@@ -568,11 +535,6 @@ variancep_ki_native_f32 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes
       }
       op_mask[out_i] = 1;
       op[out_i++] = (double) 0.0;
-    } else if ( n_valid < 1 ) {
-      /* Legacy degenerate case (e.g. sample variance with 1 unmasked
-         cell): return zero rather than UNDEF, matching pre-migration
-         behaviour of `cnt > 1 ? formula : 0`. */
-      op[out_i++] = (double) 0.0;
     } else {
       double mean = sum / (double) n_valid;
       /* Pass 2: M2 = Σ (x - mean)² */
@@ -610,7 +572,7 @@ variancep_ki_native_f64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes
                                  slab_axes, naxes, 0);
   if ( rc != CA_ITER_OK ) {
     rb_raise(rb_eRuntimeError,
-             "variancep_ki: kernel_iterator init failed rc=%d", rc);
+             "%s: kernel_iterator init failed rc=%d", ca_calling_method_name(), rc);
   }
   char       *p;
   boolean8_t *m;
@@ -622,7 +584,7 @@ variancep_ki_native_f64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes
     ca_size_t masked_cnt = 0;
     CA_SLAB_REDUCE_SUM8_EX(double, double, st, p, m, sum, 0.0, __variancep_f64_P1EXPR, masked_cnt);
     ca_size_t n_valid = st.slab_elements - masked_cnt;
-    int __trigger = (min_count < 0 ? masked_cnt == st.slab_elements : n_valid < min_count);
+    int __trigger = ca_too_few_present(n_valid, min_count, 1);
     if ( __trigger ) {
       if ( ! op_mask ) {
         ca_create_mask(co);
@@ -630,11 +592,6 @@ variancep_ki_native_f64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes
       }
       op_mask[out_i] = 1;
       op[out_i++] = (double) 0.0;
-    } else if ( n_valid < 1 ) {
-      /* Legacy degenerate case (e.g. sample variance with 1 unmasked
-         cell): return zero rather than UNDEF, matching pre-migration
-         behaviour of `cnt > 1 ? formula : 0`. */
-      op[out_i++] = (double) 0.0;
     } else {
       double mean = sum / (double) n_valid;
       /* Pass 2: M2 = Σ (x - mean)² */
@@ -672,7 +629,7 @@ variancep_ki_native_cmplx64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t n
                                  slab_axes, naxes, 0);
   if ( rc != CA_ITER_OK ) {
     rb_raise(rb_eRuntimeError,
-             "variancep_ki: kernel_iterator init failed rc=%d", rc);
+             "%s: kernel_iterator init failed rc=%d", ca_calling_method_name(), rc);
   }
   char       *p;
   boolean8_t *m;
@@ -684,7 +641,7 @@ variancep_ki_native_cmplx64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t n
     ca_size_t masked_cnt = 0;
     CA_SLAB_REDUCE_SUM8_EX(cmplx64_t, cmplx64_t, st, p, m, sum, 0, __variancep_cmplx64_P1EXPR, masked_cnt);
     ca_size_t n_valid = st.slab_elements - masked_cnt;
-    int __trigger = (min_count < 0 ? masked_cnt == st.slab_elements : n_valid < min_count);
+    int __trigger = ca_too_few_present(n_valid, min_count, 1);
     if ( __trigger ) {
       if ( ! op_mask ) {
         ca_create_mask(co);
@@ -692,11 +649,6 @@ variancep_ki_native_cmplx64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t n
       }
       op_mask[out_i] = 1;
       op[out_i++] = (double) 0.0;
-    } else if ( n_valid < 1 ) {
-      /* Legacy degenerate case (e.g. sample variance with 1 unmasked
-         cell): return zero rather than UNDEF, matching pre-migration
-         behaviour of `cnt > 1 ? formula : 0`. */
-      op[out_i++] = (double) 0.0;
     } else {
       cmplx64_t mean = sum / (cmplx64_t)(double) n_valid;
       /* Pass 2: M2 = Σ (x - mean)² */
@@ -734,7 +686,7 @@ variancep_ki_native_cmplx128 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t
                                  slab_axes, naxes, 0);
   if ( rc != CA_ITER_OK ) {
     rb_raise(rb_eRuntimeError,
-             "variancep_ki: kernel_iterator init failed rc=%d", rc);
+             "%s: kernel_iterator init failed rc=%d", ca_calling_method_name(), rc);
   }
   char       *p;
   boolean8_t *m;
@@ -746,7 +698,7 @@ variancep_ki_native_cmplx128 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t
     ca_size_t masked_cnt = 0;
     CA_SLAB_REDUCE_SUM8_EX(cmplx128_t, cmplx128_t, st, p, m, sum, 0, __variancep_cmplx128_P1EXPR, masked_cnt);
     ca_size_t n_valid = st.slab_elements - masked_cnt;
-    int __trigger = (min_count < 0 ? masked_cnt == st.slab_elements : n_valid < min_count);
+    int __trigger = ca_too_few_present(n_valid, min_count, 1);
     if ( __trigger ) {
       if ( ! op_mask ) {
         ca_create_mask(co);
@@ -754,11 +706,6 @@ variancep_ki_native_cmplx128 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t
       }
       op_mask[out_i] = 1;
       op[out_i++] = (double) 0.0;
-    } else if ( n_valid < 1 ) {
-      /* Legacy degenerate case (e.g. sample variance with 1 unmasked
-         cell): return zero rather than UNDEF, matching pre-migration
-         behaviour of `cnt > 1 ? formula : 0`. */
-      op[out_i++] = (double) 0.0;
     } else {
       cmplx128_t mean = sum / (cmplx128_t)(double) n_valid;
       /* Pass 2: M2 = Σ (x - mean)² */
@@ -796,7 +743,7 @@ variancep_ki_native_bool (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxe
                                  slab_axes, naxes, 0);
   if ( rc != CA_ITER_OK ) {
     rb_raise(rb_eRuntimeError,
-             "variancep_ki: kernel_iterator init failed rc=%d", rc);
+             "%s: kernel_iterator init failed rc=%d", ca_calling_method_name(), rc);
   }
   char       *p;
   boolean8_t *m;
@@ -808,7 +755,7 @@ variancep_ki_native_bool (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxe
     ca_size_t masked_cnt = 0;
     CA_SLAB_REDUCE_SUM8_EX(boolean8_t, double, st, p, m, sum, 0.0, __variancep_bool_P1EXPR, masked_cnt);
     ca_size_t n_valid = st.slab_elements - masked_cnt;
-    int __trigger = (min_count < 0 ? masked_cnt == st.slab_elements : n_valid < min_count);
+    int __trigger = ca_too_few_present(n_valid, min_count, 1);
     if ( __trigger ) {
       if ( ! op_mask ) {
         ca_create_mask(co);
@@ -816,11 +763,6 @@ variancep_ki_native_bool (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxe
       }
       op_mask[out_i] = 1;
       op[out_i++] = (double) 0.0;
-    } else if ( n_valid < 1 ) {
-      /* Legacy degenerate case (e.g. sample variance with 1 unmasked
-         cell): return zero rather than UNDEF, matching pre-migration
-         behaviour of `cnt > 1 ? formula : 0`. */
-      op[out_i++] = (double) 0.0;
     } else {
       double mean = sum / (double) n_valid;
       /* Pass 2: M2 = Σ (x - mean)² */
@@ -844,19 +786,45 @@ variancep_ki_native_bool (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxe
 #undef __variancep_bool_P1EXPR
 #undef __variancep_bool_P2EXPR
+/* The object lane calls Ruby for every cell, so a walk can raise part
+   way.  Its states live in the caller's frame, and ca_iter_ensure
+   finishes them however the body leaves. */
+typedef struct {
+  ca_iter_state states[1];
+  VALUE self;
+  CArray *ca;
+  int8_t *slab_axes;
+  int8_t naxes;
+  int keep_axis;
+  ca_size_t min_count;
+} variancep_ki_native_object_frame;
 static VALUE
-variancep_ki_native_object (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int keep_axis, ca_size_t min_count)
+variancep_ki_native_object_body (VALUE arg)
 {
+  variancep_ki_native_object_frame *fr = (variancep_ki_native_object_frame *) arg;
+  VALUE self = fr->self;
+  (void) self;
+  CArray *ca = fr->ca;
+  (void) ca;
+  int8_t *slab_axes = fr->slab_axes;
+  (void) slab_axes;
+  int8_t naxes = fr->naxes;
+  (void) naxes;
+  int keep_axis = fr->keep_axis;
+  (void) keep_axis;
+  ca_size_t min_count = fr->min_count;
+  (void) min_count;
+#define st (fr->states[0])
   VALUE   vout  = rb_ca_new_reduced(self, slab_axes, naxes, CA_OBJECT, keep_axis);
   CArray *co;
   GetCArray(vout, co);
   VALUE *op = (VALUE *) co->ptr;
-  ca_iter_state st;
   int rc = ca_iter_state_init_l2(&st, ca, CA_SLAB_AXES,
                                  slab_axes, naxes, 0);
   if ( rc != CA_ITER_OK ) {
     rb_raise(rb_eRuntimeError,
-             "variancep_ki: kernel_iterator init failed rc=%d", rc);
+             "%s: kernel_iterator init failed rc=%d", ca_calling_method_name(), rc);
   }
   char       *p;
   boolean8_t *m;
@@ -868,7 +836,7 @@ variancep_ki_native_object (VALUE self, CArray *ca, int8_t *slab_axes, int8_t na
     ca_size_t masked_cnt = 0;
     CA_SLAB_REDUCE_T_EX(VALUE, st, p, m, sum, INT2FIX(0), sum = rb_funcall(sum, rb_intern("+"), 1, v), masked_cnt);
     ca_size_t n_valid = st.slab_elements - masked_cnt;
-    int __trigger = (min_count < 0 ? masked_cnt == st.slab_elements : n_valid < min_count);
+    int __trigger = ca_too_few_present(n_valid, min_count, 1);
     if ( __trigger ) {
       if ( ! op_mask ) {
         ca_create_mask(co);
@@ -876,11 +844,6 @@ variancep_ki_native_object (VALUE self, CArray *ca, int8_t *slab_axes, int8_t na
       }
       op_mask[out_i] = 1;
       op[out_i++] = (VALUE) INT2FIX(0);
-    } else if ( n_valid < 1 ) {
-      /* Legacy degenerate case (e.g. sample variance with 1 unmasked
-         cell): return zero rather than UNDEF, matching pre-migration
-         behaviour of `cnt > 1 ? formula : 0`. */
-      op[out_i++] = (VALUE) INT2FIX(0);
     } else {
       VALUE mean = rb_funcall(sum, rb_intern("/"), 1, LONG2NUM(n_valid));
       /* Pass 2: M2 = Σ (x - mean)² */
@@ -901,8 +864,23 @@ M2 = rb_funcall(M2, rb_intern("+"), 1, rb_funcall(__d, rb_intern("*"), 1, __d)),
     return (VALUE)(op[0]);
   }
   return vout;
+#undef st
 }
+static VALUE
+variancep_ki_native_object (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int keep_axis, ca_size_t min_count)
+{
+  variancep_ki_native_object_frame fr;
+  fr.self = self;
+  fr.ca = ca;
+  fr.slab_axes = slab_axes;
+  fr.naxes = naxes;
+  fr.keep_axis = keep_axis;
+  fr.min_count = min_count;
+  return ca_iter_ensure(1, fr.states, variancep_ki_native_object_body, (VALUE) &fr);
+}
 VALUE
 rb_ca_variancep_ki (int argc, VALUE *argv, VALUE self)
 {
@@ -914,89 +892,101 @@ rb_ca_variancep_ki (int argc, VALUE *argv, VALUE self)
   int keep_axis = RTEST(rkeep_axis);
   int8_t slab_axes[CA_RANK_MAX];
   int8_t naxes = rb_ca_parse_reduce_axes_kw(raxis, src, slab_axes);
-  ca_size_t min_count = -1;
-  if ( ! NIL_P(rmin_count) ) {
-    ca_size_t mc_user = NUM2SIZE(rmin_count);
-    if ( mc_user < 0 ) {
-      rb_raise(rb_eArgError, "variancep_ki: :min_count must be non-negative, got %lld", (long long) mc_user);
+  if ( ca_is_lazy_view(src) ) {
+    volatile VALUE evaluated = ca_lazy_evaluated(self);
+    if ( ! NIL_P(evaluated) ) {
+      self = evaluated;
+      GetCArray(self, src);
     }
-    /* :min_count = 0 collapses to legacy default (any 1 valid). */
-    if ( mc_user > 0 ) min_count = mc_user;
+  }
+  ca_size_t min_count = ca_min_count_arg(rmin_count, NULL);
+  if ( rfval == CA_UNDEF ) rfval = Qnil;
+  int fill_whole = 0;
+  if ( ! NIL_P(rfval) && naxes == src->ndim && ! keep_axis ) {
+    keep_axis  = 1;
+    fill_whole = 1;
   }
   VALUE result;
   if ( argc > 0 ) {
-    rb_raise(rb_eArgError, "variancep_ki: positional axis arguments are no longer accepted (got %d); use axis: kwarg, e.g. a.variancep(axis: 0) or a.variancep(axis: [0, 1])", argc);
+    rb_raise(rb_eArgError, "%s: positional axis arguments are no longer accepted (got %d); use the axis: keyword, e.g. a.%s(axis: 0) or a.%s(axis: [0, 1])", ca_calling_method_name(), argc, ca_calling_method_name(), ca_calling_method_name());
   }
   switch ( src->data_type ) {
     case CA_INT8: {
       result = variancep_ki_native_i8(self, src, slab_axes, naxes, keep_axis, min_count);
+      RB_GC_GUARD(self);
       break;
     }
     case CA_UINT8: {
       result = variancep_ki_native_u8(self, src, slab_axes, naxes, keep_axis, min_count);
+      RB_GC_GUARD(self);
       break;
     }
     case CA_INT16: {
       result = variancep_ki_native_i16(self, src, slab_axes, naxes, keep_axis, min_count);
+      RB_GC_GUARD(self);
       break;
     }
     case CA_UINT16: {
       result = variancep_ki_native_u16(self, src, slab_axes, naxes, keep_axis, min_count);
+      RB_GC_GUARD(self);
       break;
     }
     case CA_INT32: {
       result = variancep_ki_native_i32(self, src, slab_axes, naxes, keep_axis, min_count);
+      RB_GC_GUARD(self);
       break;
     }
     case CA_UINT32: {
       result = variancep_ki_native_u32(self, src, slab_axes, naxes, keep_axis, min_count);
+      RB_GC_GUARD(self);
       break;
     }
     case CA_INT64: {
       result = variancep_ki_native_i64(self, src, slab_axes, naxes, keep_axis, min_count);
+      RB_GC_GUARD(self);
       break;
     }
     case CA_UINT64: {
       result = variancep_ki_native_u64(self, src, slab_axes, naxes, keep_axis, min_count);
+      RB_GC_GUARD(self);
       break;
     }
     case CA_FLOAT32: {
       result = variancep_ki_native_f32(self, src, slab_axes, naxes, keep_axis, min_count);
+      RB_GC_GUARD(self);
       break;
     }
     case CA_FLOAT64: {
       result = variancep_ki_native_f64(self, src, slab_axes, naxes, keep_axis, min_count);
+      RB_GC_GUARD(self);
       break;
     }
     case CA_CMPLX64: {
       result = variancep_ki_native_cmplx64(self, src, slab_axes, naxes, keep_axis, min_count);
+      RB_GC_GUARD(self);
       break;
     }
     case CA_CMPLX128: {
       result = variancep_ki_native_cmplx128(self, src, slab_axes, naxes, keep_axis, min_count);
+      RB_GC_GUARD(self);
       break;
     }
     case CA_BOOLEAN: {
       result = variancep_ki_native_bool(self, src, slab_axes, naxes, keep_axis, min_count);
+      RB_GC_GUARD(self);
       break;
     }
     case CA_OBJECT: {
       result = variancep_ki_native_object(self, src, slab_axes, naxes, keep_axis, min_count);
+      RB_GC_GUARD(self);
       break;
     }
     default:
-      rb_raise(rb_eCADataTypeError, "variancep_ki: source data_type :%s not supported (expected one of: i8, u8, i16, u16, i32, u32, i64, u64, f32, f64, cmplx64, cmplx128, bool, object)", ca_type_name[src->data_type]);
+      ca_face_refuse_numeric(self, ca_calling_method_name(), rb_eCADataTypeError);
+      rb_raise(rb_eCADataTypeError, "%s: source data_type :%s not supported (expected one of: int8, uint8, int16, uint16, int32, uint32, int64, uint64, float32, float64, cmplx64, cmplx128, boolean, object)", ca_calling_method_name(), ca_type_name[src->data_type]);
   }
   if ( ! NIL_P(rfval) ) {
-    if ( result == CA_UNDEF ) {
-      result = rfval;
-    } else if ( TYPE(result) != T_FLOAT && rb_obj_is_kind_of(result, rb_cCArray) ) {
-      CArray *cr;
-      GetCArray(result, cr);
-      if ( ca_has_mask(cr) ) {
-        result = rb_ca_mask_fill_copy(result, rfval);
-      }
-    }
+    result = ca_reduce_fill(result, rfval, fill_whole);
   }
   return result;
 }
@@ -1017,7 +1007,7 @@ variance_ki_native_i8 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes,
                                  slab_axes, naxes, 0);
   if ( rc != CA_ITER_OK ) {
     rb_raise(rb_eRuntimeError,
-             "variance_ki: kernel_iterator init failed rc=%d", rc);
+             "%s: kernel_iterator init failed rc=%d", ca_calling_method_name(), rc);
   }
   char       *p;
   boolean8_t *m;
@@ -1029,7 +1019,7 @@ variance_ki_native_i8 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes,
     ca_size_t masked_cnt = 0;
     CA_SLAB_REDUCE_SUM8_EX(int8_t, double, st, p, m, sum, 0.0, __variance_i8_P1EXPR, masked_cnt);
     ca_size_t n_valid = st.slab_elements - masked_cnt;
-    int __trigger = (min_count < 0 ? masked_cnt == st.slab_elements : n_valid < min_count);
+    int __trigger = ca_too_few_present(n_valid, min_count, 2);
     if ( __trigger ) {
       if ( ! op_mask ) {
         ca_create_mask(co);
@@ -1037,11 +1027,6 @@ variance_ki_native_i8 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes,
       }
       op_mask[out_i] = 1;
       op[out_i++] = (double) 0.0;
-    } else if ( n_valid < 2 ) {
-      /* Legacy degenerate case (e.g. sample variance with 1 unmasked
-         cell): return zero rather than UNDEF, matching pre-migration
-         behaviour of `cnt > 1 ? formula : 0`. */
-      op[out_i++] = (double) 0.0;
     } else {
       double mean = sum / (double) n_valid;
       /* Pass 2: M2 = Σ (x - mean)² */
@@ -1079,7 +1064,7 @@ variance_ki_native_u8 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes,
                                  slab_axes, naxes, 0);
   if ( rc != CA_ITER_OK ) {
     rb_raise(rb_eRuntimeError,
-             "variance_ki: kernel_iterator init failed rc=%d", rc);
+             "%s: kernel_iterator init failed rc=%d", ca_calling_method_name(), rc);
   }
   char       *p;
   boolean8_t *m;
@@ -1091,7 +1076,7 @@ variance_ki_native_u8 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes,
     ca_size_t masked_cnt = 0;
     CA_SLAB_REDUCE_SUM8_EX(uint8_t, double, st, p, m, sum, 0.0, __variance_u8_P1EXPR, masked_cnt);
     ca_size_t n_valid = st.slab_elements - masked_cnt;
-    int __trigger = (min_count < 0 ? masked_cnt == st.slab_elements : n_valid < min_count);
+    int __trigger = ca_too_few_present(n_valid, min_count, 2);
     if ( __trigger ) {
       if ( ! op_mask ) {
         ca_create_mask(co);
@@ -1099,11 +1084,6 @@ variance_ki_native_u8 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes,
       }
       op_mask[out_i] = 1;
       op[out_i++] = (double) 0.0;
-    } else if ( n_valid < 2 ) {
-      /* Legacy degenerate case (e.g. sample variance with 1 unmasked
-         cell): return zero rather than UNDEF, matching pre-migration
-         behaviour of `cnt > 1 ? formula : 0`. */
-      op[out_i++] = (double) 0.0;
     } else {
       double mean = sum / (double) n_valid;
       /* Pass 2: M2 = Σ (x - mean)² */
@@ -1141,7 +1121,7 @@ variance_ki_native_i16 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes,
                                  slab_axes, naxes, 0);
   if ( rc != CA_ITER_OK ) {
     rb_raise(rb_eRuntimeError,
-             "variance_ki: kernel_iterator init failed rc=%d", rc);
+             "%s: kernel_iterator init failed rc=%d", ca_calling_method_name(), rc);
   }
   char       *p;
   boolean8_t *m;
@@ -1153,7 +1133,7 @@ variance_ki_native_i16 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes,
     ca_size_t masked_cnt = 0;
     CA_SLAB_REDUCE_SUM8_EX(int16_t, double, st, p, m, sum, 0.0, __variance_i16_P1EXPR, masked_cnt);
     ca_size_t n_valid = st.slab_elements - masked_cnt;
-    int __trigger = (min_count < 0 ? masked_cnt == st.slab_elements : n_valid < min_count);
+    int __trigger = ca_too_few_present(n_valid, min_count, 2);
     if ( __trigger ) {
       if ( ! op_mask ) {
         ca_create_mask(co);
@@ -1161,11 +1141,6 @@ variance_ki_native_i16 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes,
       }
       op_mask[out_i] = 1;
       op[out_i++] = (double) 0.0;
-    } else if ( n_valid < 2 ) {
-      /* Legacy degenerate case (e.g. sample variance with 1 unmasked
-         cell): return zero rather than UNDEF, matching pre-migration
-         behaviour of `cnt > 1 ? formula : 0`. */
-      op[out_i++] = (double) 0.0;
     } else {
       double mean = sum / (double) n_valid;
       /* Pass 2: M2 = Σ (x - mean)² */
@@ -1203,7 +1178,7 @@ variance_ki_native_u16 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes,
                                  slab_axes, naxes, 0);
   if ( rc != CA_ITER_OK ) {
     rb_raise(rb_eRuntimeError,
-             "variance_ki: kernel_iterator init failed rc=%d", rc);
+             "%s: kernel_iterator init failed rc=%d", ca_calling_method_name(), rc);
   }
   char       *p;
   boolean8_t *m;
@@ -1215,7 +1190,7 @@ variance_ki_native_u16 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes,
     ca_size_t masked_cnt = 0;
     CA_SLAB_REDUCE_SUM8_EX(uint16_t, double, st, p, m, sum, 0.0, __variance_u16_P1EXPR, masked_cnt);
     ca_size_t n_valid = st.slab_elements - masked_cnt;
-    int __trigger = (min_count < 0 ? masked_cnt == st.slab_elements : n_valid < min_count);
+    int __trigger = ca_too_few_present(n_valid, min_count, 2);
     if ( __trigger ) {
       if ( ! op_mask ) {
         ca_create_mask(co);
@@ -1223,11 +1198,6 @@ variance_ki_native_u16 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes,
       }
       op_mask[out_i] = 1;
       op[out_i++] = (double) 0.0;
-    } else if ( n_valid < 2 ) {
-      /* Legacy degenerate case (e.g. sample variance with 1 unmasked
-         cell): return zero rather than UNDEF, matching pre-migration
-         behaviour of `cnt > 1 ? formula : 0`. */
-      op[out_i++] = (double) 0.0;
     } else {
       double mean = sum / (double) n_valid;
       /* Pass 2: M2 = Σ (x - mean)² */
@@ -1265,7 +1235,7 @@ variance_ki_native_i32 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes,
                                  slab_axes, naxes, 0);
   if ( rc != CA_ITER_OK ) {
     rb_raise(rb_eRuntimeError,
-             "variance_ki: kernel_iterator init failed rc=%d", rc);
+             "%s: kernel_iterator init failed rc=%d", ca_calling_method_name(), rc);
   }
   char       *p;
   boolean8_t *m;
@@ -1277,7 +1247,7 @@ variance_ki_native_i32 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes,
     ca_size_t masked_cnt = 0;
     CA_SLAB_REDUCE_SUM8_EX(int32_t, double, st, p, m, sum, 0.0, __variance_i32_P1EXPR, masked_cnt);
     ca_size_t n_valid = st.slab_elements - masked_cnt;
-    int __trigger = (min_count < 0 ? masked_cnt == st.slab_elements : n_valid < min_count);
+    int __trigger = ca_too_few_present(n_valid, min_count, 2);
     if ( __trigger ) {
       if ( ! op_mask ) {
         ca_create_mask(co);
@@ -1285,11 +1255,6 @@ variance_ki_native_i32 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes,
       }
       op_mask[out_i] = 1;
       op[out_i++] = (double) 0.0;
-    } else if ( n_valid < 2 ) {
-      /* Legacy degenerate case (e.g. sample variance with 1 unmasked
-         cell): return zero rather than UNDEF, matching pre-migration
-         behaviour of `cnt > 1 ? formula : 0`. */
-      op[out_i++] = (double) 0.0;
     } else {
       double mean = sum / (double) n_valid;
       /* Pass 2: M2 = Σ (x - mean)² */
@@ -1327,7 +1292,7 @@ variance_ki_native_u32 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes,
                                  slab_axes, naxes, 0);
   if ( rc != CA_ITER_OK ) {
     rb_raise(rb_eRuntimeError,
-             "variance_ki: kernel_iterator init failed rc=%d", rc);
+             "%s: kernel_iterator init failed rc=%d", ca_calling_method_name(), rc);
   }
   char       *p;
   boolean8_t *m;
@@ -1339,7 +1304,7 @@ variance_ki_native_u32 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes,
     ca_size_t masked_cnt = 0;
     CA_SLAB_REDUCE_SUM8_EX(uint32_t, double, st, p, m, sum, 0.0, __variance_u32_P1EXPR, masked_cnt);
     ca_size_t n_valid = st.slab_elements - masked_cnt;
-    int __trigger = (min_count < 0 ? masked_cnt == st.slab_elements : n_valid < min_count);
+    int __trigger = ca_too_few_present(n_valid, min_count, 2);
     if ( __trigger ) {
       if ( ! op_mask ) {
         ca_create_mask(co);
@@ -1347,11 +1312,6 @@ variance_ki_native_u32 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes,
       }
       op_mask[out_i] = 1;
       op[out_i++] = (double) 0.0;
-    } else if ( n_valid < 2 ) {
-      /* Legacy degenerate case (e.g. sample variance with 1 unmasked
-         cell): return zero rather than UNDEF, matching pre-migration
-         behaviour of `cnt > 1 ? formula : 0`. */
-      op[out_i++] = (double) 0.0;
     } else {
       double mean = sum / (double) n_valid;
       /* Pass 2: M2 = Σ (x - mean)² */
@@ -1389,7 +1349,7 @@ variance_ki_native_i64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes,
                                  slab_axes, naxes, 0);
   if ( rc != CA_ITER_OK ) {
     rb_raise(rb_eRuntimeError,
-             "variance_ki: kernel_iterator init failed rc=%d", rc);
+             "%s: kernel_iterator init failed rc=%d", ca_calling_method_name(), rc);
   }
   char       *p;
   boolean8_t *m;
@@ -1401,7 +1361,7 @@ variance_ki_native_i64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes,
     ca_size_t masked_cnt = 0;
     CA_SLAB_REDUCE_SUM8_EX(int64_t, double, st, p, m, sum, 0.0, __variance_i64_P1EXPR, masked_cnt);
     ca_size_t n_valid = st.slab_elements - masked_cnt;
-    int __trigger = (min_count < 0 ? masked_cnt == st.slab_elements : n_valid < min_count);
+    int __trigger = ca_too_few_present(n_valid, min_count, 2);
     if ( __trigger ) {
       if ( ! op_mask ) {
         ca_create_mask(co);
@@ -1409,11 +1369,6 @@ variance_ki_native_i64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes,
       }
       op_mask[out_i] = 1;
       op[out_i++] = (double) 0.0;
-    } else if ( n_valid < 2 ) {
-      /* Legacy degenerate case (e.g. sample variance with 1 unmasked
-         cell): return zero rather than UNDEF, matching pre-migration
-         behaviour of `cnt > 1 ? formula : 0`. */
-      op[out_i++] = (double) 0.0;
     } else {
       double mean = sum / (double) n_valid;
       /* Pass 2: M2 = Σ (x - mean)² */
@@ -1451,7 +1406,7 @@ variance_ki_native_u64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes,
                                  slab_axes, naxes, 0);
   if ( rc != CA_ITER_OK ) {
     rb_raise(rb_eRuntimeError,
-             "variance_ki: kernel_iterator init failed rc=%d", rc);
+             "%s: kernel_iterator init failed rc=%d", ca_calling_method_name(), rc);
   }
   char       *p;
   boolean8_t *m;
@@ -1463,7 +1418,7 @@ variance_ki_native_u64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes,
     ca_size_t masked_cnt = 0;
     CA_SLAB_REDUCE_SUM8_EX(uint64_t, double, st, p, m, sum, 0.0, __variance_u64_P1EXPR, masked_cnt);
     ca_size_t n_valid = st.slab_elements - masked_cnt;
-    int __trigger = (min_count < 0 ? masked_cnt == st.slab_elements : n_valid < min_count);
+    int __trigger = ca_too_few_present(n_valid, min_count, 2);
     if ( __trigger ) {
       if ( ! op_mask ) {
         ca_create_mask(co);
@@ -1471,11 +1426,6 @@ variance_ki_native_u64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes,
       }
       op_mask[out_i] = 1;
       op[out_i++] = (double) 0.0;
-    } else if ( n_valid < 2 ) {
-      /* Legacy degenerate case (e.g. sample variance with 1 unmasked
-         cell): return zero rather than UNDEF, matching pre-migration
-         behaviour of `cnt > 1 ? formula : 0`. */
-      op[out_i++] = (double) 0.0;
     } else {
       double mean = sum / (double) n_valid;
       /* Pass 2: M2 = Σ (x - mean)² */
@@ -1513,7 +1463,7 @@ variance_ki_native_f32 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes,
                                  slab_axes, naxes, 0);
   if ( rc != CA_ITER_OK ) {
     rb_raise(rb_eRuntimeError,
-             "variance_ki: kernel_iterator init failed rc=%d", rc);
+             "%s: kernel_iterator init failed rc=%d", ca_calling_method_name(), rc);
   }
   char       *p;
   boolean8_t *m;
@@ -1525,7 +1475,7 @@ variance_ki_native_f32 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes,
     ca_size_t masked_cnt = 0;
     CA_SLAB_REDUCE_SUM8_EX(float, double, st, p, m, sum, 0.0, __variance_f32_P1EXPR, masked_cnt);
     ca_size_t n_valid = st.slab_elements - masked_cnt;
-    int __trigger = (min_count < 0 ? masked_cnt == st.slab_elements : n_valid < min_count);
+    int __trigger = ca_too_few_present(n_valid, min_count, 2);
     if ( __trigger ) {
       if ( ! op_mask ) {
         ca_create_mask(co);
@@ -1533,11 +1483,6 @@ variance_ki_native_f32 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes,
       }
       op_mask[out_i] = 1;
       op[out_i++] = (double) 0.0;
-    } else if ( n_valid < 2 ) {
-      /* Legacy degenerate case (e.g. sample variance with 1 unmasked
-         cell): return zero rather than UNDEF, matching pre-migration
-         behaviour of `cnt > 1 ? formula : 0`. */
-      op[out_i++] = (double) 0.0;
     } else {
       double mean = sum / (double) n_valid;
       /* Pass 2: M2 = Σ (x - mean)² */
@@ -1575,7 +1520,7 @@ variance_ki_native_f64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes,
                                  slab_axes, naxes, 0);
   if ( rc != CA_ITER_OK ) {
     rb_raise(rb_eRuntimeError,
-             "variance_ki: kernel_iterator init failed rc=%d", rc);
+             "%s: kernel_iterator init failed rc=%d", ca_calling_method_name(), rc);
   }
   char       *p;
   boolean8_t *m;
@@ -1587,7 +1532,7 @@ variance_ki_native_f64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes,
     ca_size_t masked_cnt = 0;
     CA_SLAB_REDUCE_SUM8_EX(double, double, st, p, m, sum, 0.0, __variance_f64_P1EXPR, masked_cnt);
     ca_size_t n_valid = st.slab_elements - masked_cnt;
-    int __trigger = (min_count < 0 ? masked_cnt == st.slab_elements : n_valid < min_count);
+    int __trigger = ca_too_few_present(n_valid, min_count, 2);
     if ( __trigger ) {
       if ( ! op_mask ) {
         ca_create_mask(co);
@@ -1595,11 +1540,6 @@ variance_ki_native_f64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes,
       }
       op_mask[out_i] = 1;
       op[out_i++] = (double) 0.0;
-    } else if ( n_valid < 2 ) {
-      /* Legacy degenerate case (e.g. sample variance with 1 unmasked
-         cell): return zero rather than UNDEF, matching pre-migration
-         behaviour of `cnt > 1 ? formula : 0`. */
-      op[out_i++] = (double) 0.0;
     } else {
       double mean = sum / (double) n_valid;
       /* Pass 2: M2 = Σ (x - mean)² */
@@ -1637,7 +1577,7 @@ variance_ki_native_cmplx64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t na
                                  slab_axes, naxes, 0);
   if ( rc != CA_ITER_OK ) {
     rb_raise(rb_eRuntimeError,
-             "variance_ki: kernel_iterator init failed rc=%d", rc);
+             "%s: kernel_iterator init failed rc=%d", ca_calling_method_name(), rc);
   }
   char       *p;
   boolean8_t *m;
@@ -1649,7 +1589,7 @@ variance_ki_native_cmplx64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t na
     ca_size_t masked_cnt = 0;
     CA_SLAB_REDUCE_SUM8_EX(cmplx64_t, cmplx64_t, st, p, m, sum, 0, __variance_cmplx64_P1EXPR, masked_cnt);
     ca_size_t n_valid = st.slab_elements - masked_cnt;
-    int __trigger = (min_count < 0 ? masked_cnt == st.slab_elements : n_valid < min_count);
+    int __trigger = ca_too_few_present(n_valid, min_count, 2);
     if ( __trigger ) {
       if ( ! op_mask ) {
         ca_create_mask(co);
@@ -1657,11 +1597,6 @@ variance_ki_native_cmplx64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t na
       }
       op_mask[out_i] = 1;
       op[out_i++] = (double) 0.0;
-    } else if ( n_valid < 2 ) {
-      /* Legacy degenerate case (e.g. sample variance with 1 unmasked
-         cell): return zero rather than UNDEF, matching pre-migration
-         behaviour of `cnt > 1 ? formula : 0`. */
-      op[out_i++] = (double) 0.0;
     } else {
       cmplx64_t mean = sum / (cmplx64_t)(double) n_valid;
       /* Pass 2: M2 = Σ (x - mean)² */
@@ -1699,7 +1634,7 @@ variance_ki_native_cmplx128 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t n
                                  slab_axes, naxes, 0);
   if ( rc != CA_ITER_OK ) {
     rb_raise(rb_eRuntimeError,
-             "variance_ki: kernel_iterator init failed rc=%d", rc);
+             "%s: kernel_iterator init failed rc=%d", ca_calling_method_name(), rc);
   }
   char       *p;
   boolean8_t *m;
@@ -1711,7 +1646,7 @@ variance_ki_native_cmplx128 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t n
     ca_size_t masked_cnt = 0;
     CA_SLAB_REDUCE_SUM8_EX(cmplx128_t, cmplx128_t, st, p, m, sum, 0, __variance_cmplx128_P1EXPR, masked_cnt);
     ca_size_t n_valid = st.slab_elements - masked_cnt;
-    int __trigger = (min_count < 0 ? masked_cnt == st.slab_elements : n_valid < min_count);
+    int __trigger = ca_too_few_present(n_valid, min_count, 2);
     if ( __trigger ) {
       if ( ! op_mask ) {
         ca_create_mask(co);
@@ -1719,11 +1654,6 @@ variance_ki_native_cmplx128 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t n
       }
       op_mask[out_i] = 1;
       op[out_i++] = (double) 0.0;
-    } else if ( n_valid < 2 ) {
-      /* Legacy degenerate case (e.g. sample variance with 1 unmasked
-         cell): return zero rather than UNDEF, matching pre-migration
-         behaviour of `cnt > 1 ? formula : 0`. */
-      op[out_i++] = (double) 0.0;
     } else {
       cmplx128_t mean = sum / (cmplx128_t)(double) n_valid;
       /* Pass 2: M2 = Σ (x - mean)² */
@@ -1761,7 +1691,7 @@ variance_ki_native_bool (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes
                                  slab_axes, naxes, 0);
   if ( rc != CA_ITER_OK ) {
     rb_raise(rb_eRuntimeError,
-             "variance_ki: kernel_iterator init failed rc=%d", rc);
+             "%s: kernel_iterator init failed rc=%d", ca_calling_method_name(), rc);
   }
   char       *p;
   boolean8_t *m;
@@ -1773,7 +1703,7 @@ variance_ki_native_bool (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes
     ca_size_t masked_cnt = 0;
     CA_SLAB_REDUCE_SUM8_EX(boolean8_t, double, st, p, m, sum, 0.0, __variance_bool_P1EXPR, masked_cnt);
     ca_size_t n_valid = st.slab_elements - masked_cnt;
-    int __trigger = (min_count < 0 ? masked_cnt == st.slab_elements : n_valid < min_count);
+    int __trigger = ca_too_few_present(n_valid, min_count, 2);
     if ( __trigger ) {
       if ( ! op_mask ) {
         ca_create_mask(co);
@@ -1781,11 +1711,6 @@ variance_ki_native_bool (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes
       }
       op_mask[out_i] = 1;
       op[out_i++] = (double) 0.0;
-    } else if ( n_valid < 2 ) {
-      /* Legacy degenerate case (e.g. sample variance with 1 unmasked
-         cell): return zero rather than UNDEF, matching pre-migration
-         behaviour of `cnt > 1 ? formula : 0`. */
-      op[out_i++] = (double) 0.0;
     } else {
       double mean = sum / (double) n_valid;
       /* Pass 2: M2 = Σ (x - mean)² */
@@ -1809,19 +1734,45 @@ variance_ki_native_bool (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes
 #undef __variance_bool_P1EXPR
 #undef __variance_bool_P2EXPR
+/* The object lane calls Ruby for every cell, so a walk can raise part
+   way.  Its states live in the caller's frame, and ca_iter_ensure
+   finishes them however the body leaves. */
+typedef struct {
+  ca_iter_state states[1];
+  VALUE self;
+  CArray *ca;
+  int8_t *slab_axes;
+  int8_t naxes;
+  int keep_axis;
+  ca_size_t min_count;
+} variance_ki_native_object_frame;
 static VALUE
-variance_ki_native_object (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int keep_axis, ca_size_t min_count)
+variance_ki_native_object_body (VALUE arg)
 {
+  variance_ki_native_object_frame *fr = (variance_ki_native_object_frame *) arg;
+  VALUE self = fr->self;
+  (void) self;
+  CArray *ca = fr->ca;
+  (void) ca;
+  int8_t *slab_axes = fr->slab_axes;
+  (void) slab_axes;
+  int8_t naxes = fr->naxes;
+  (void) naxes;
+  int keep_axis = fr->keep_axis;
+  (void) keep_axis;
+  ca_size_t min_count = fr->min_count;
+  (void) min_count;
+#define st (fr->states[0])
   VALUE   vout  = rb_ca_new_reduced(self, slab_axes, naxes, CA_OBJECT, keep_axis);
   CArray *co;
   GetCArray(vout, co);
   VALUE *op = (VALUE *) co->ptr;
-  ca_iter_state st;
   int rc = ca_iter_state_init_l2(&st, ca, CA_SLAB_AXES,
                                  slab_axes, naxes, 0);
   if ( rc != CA_ITER_OK ) {
     rb_raise(rb_eRuntimeError,
-             "variance_ki: kernel_iterator init failed rc=%d", rc);
+             "%s: kernel_iterator init failed rc=%d", ca_calling_method_name(), rc);
   }
   char       *p;
   boolean8_t *m;
@@ -1833,7 +1784,7 @@ variance_ki_native_object (VALUE self, CArray *ca, int8_t *slab_axes, int8_t nax
     ca_size_t masked_cnt = 0;
     CA_SLAB_REDUCE_T_EX(VALUE, st, p, m, sum, INT2FIX(0), sum = rb_funcall(sum, rb_intern("+"), 1, v), masked_cnt);
     ca_size_t n_valid = st.slab_elements - masked_cnt;
-    int __trigger = (min_count < 0 ? masked_cnt == st.slab_elements : n_valid < min_count);
+    int __trigger = ca_too_few_present(n_valid, min_count, 2);
     if ( __trigger ) {
       if ( ! op_mask ) {
         ca_create_mask(co);
@@ -1841,11 +1792,6 @@ variance_ki_native_object (VALUE self, CArray *ca, int8_t *slab_axes, int8_t nax
       }
       op_mask[out_i] = 1;
       op[out_i++] = (VALUE) INT2FIX(0);
-    } else if ( n_valid < 2 ) {
-      /* Legacy degenerate case (e.g. sample variance with 1 unmasked
-         cell): return zero rather than UNDEF, matching pre-migration
-         behaviour of `cnt > 1 ? formula : 0`. */
-      op[out_i++] = (VALUE) INT2FIX(0);
     } else {
       VALUE mean = rb_funcall(sum, rb_intern("/"), 1, LONG2NUM(n_valid));
       /* Pass 2: M2 = Σ (x - mean)² */
@@ -1866,8 +1812,23 @@ M2 = rb_funcall(M2, rb_intern("+"), 1, rb_funcall(__d, rb_intern("*"), 1, __d)),
     return (VALUE)(op[0]);
   }
   return vout;
+#undef st
 }
+static VALUE
+variance_ki_native_object (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int keep_axis, ca_size_t min_count)
+{
+  variance_ki_native_object_frame fr;
+  fr.self = self;
+  fr.ca = ca;
+  fr.slab_axes = slab_axes;
+  fr.naxes = naxes;
+  fr.keep_axis = keep_axis;
+  fr.min_count = min_count;
+  return ca_iter_ensure(1, fr.states, variance_ki_native_object_body, (VALUE) &fr);
+}
 VALUE
 rb_ca_variance_ki (int argc, VALUE *argv, VALUE self)
 {
@@ -1879,89 +1840,101 @@ rb_ca_variance_ki (int argc, VALUE *argv, VALUE self)
   int keep_axis = RTEST(rkeep_axis);
   int8_t slab_axes[CA_RANK_MAX];
   int8_t naxes = rb_ca_parse_reduce_axes_kw(raxis, src, slab_axes);
-  ca_size_t min_count = -1;
-  if ( ! NIL_P(rmin_count) ) {
-    ca_size_t mc_user = NUM2SIZE(rmin_count);
-    if ( mc_user < 0 ) {
-      rb_raise(rb_eArgError, "variance_ki: :min_count must be non-negative, got %lld", (long long) mc_user);
+  if ( ca_is_lazy_view(src) ) {
+    volatile VALUE evaluated = ca_lazy_evaluated(self);
+    if ( ! NIL_P(evaluated) ) {
+      self = evaluated;
+      GetCArray(self, src);
     }
-    /* :min_count = 0 collapses to legacy default (any 1 valid). */
-    if ( mc_user > 0 ) min_count = mc_user;
+  }
+  ca_size_t min_count = ca_min_count_arg(rmin_count, NULL);
+  if ( rfval == CA_UNDEF ) rfval = Qnil;
+  int fill_whole = 0;
+  if ( ! NIL_P(rfval) && naxes == src->ndim && ! keep_axis ) {
+    keep_axis  = 1;
+    fill_whole = 1;
   }
   VALUE result;
   if ( argc > 0 ) {
-    rb_raise(rb_eArgError, "variance_ki: positional axis arguments are no longer accepted (got %d); use axis: kwarg, e.g. a.variance(axis: 0) or a.variance(axis: [0, 1])", argc);
+    rb_raise(rb_eArgError, "%s: positional axis arguments are no longer accepted (got %d); use the axis: keyword, e.g. a.%s(axis: 0) or a.%s(axis: [0, 1])", ca_calling_method_name(), argc, ca_calling_method_name(), ca_calling_method_name());
   }
   switch ( src->data_type ) {
     case CA_INT8: {
       result = variance_ki_native_i8(self, src, slab_axes, naxes, keep_axis, min_count);
+      RB_GC_GUARD(self);
       break;
     }
     case CA_UINT8: {
       result = variance_ki_native_u8(self, src, slab_axes, naxes, keep_axis, min_count);
+      RB_GC_GUARD(self);
       break;
     }
     case CA_INT16: {
       result = variance_ki_native_i16(self, src, slab_axes, naxes, keep_axis, min_count);
+      RB_GC_GUARD(self);
       break;
     }
     case CA_UINT16: {
       result = variance_ki_native_u16(self, src, slab_axes, naxes, keep_axis, min_count);
+      RB_GC_GUARD(self);
       break;
     }
     case CA_INT32: {
       result = variance_ki_native_i32(self, src, slab_axes, naxes, keep_axis, min_count);
+      RB_GC_GUARD(self);
       break;
     }
     case CA_UINT32: {
       result = variance_ki_native_u32(self, src, slab_axes, naxes, keep_axis, min_count);
+      RB_GC_GUARD(self);
       break;
     }
     case CA_INT64: {
       result = variance_ki_native_i64(self, src, slab_axes, naxes, keep_axis, min_count);
+      RB_GC_GUARD(self);
       break;
     }
     case CA_UINT64: {
       result = variance_ki_native_u64(self, src, slab_axes, naxes, keep_axis, min_count);
+      RB_GC_GUARD(self);
       break;
     }
     case CA_FLOAT32: {
       result = variance_ki_native_f32(self, src, slab_axes, naxes, keep_axis, min_count);
+      RB_GC_GUARD(self);
       break;
     }
     case CA_FLOAT64: {
       result = variance_ki_native_f64(self, src, slab_axes, naxes, keep_axis, min_count);
+      RB_GC_GUARD(self);
       break;
     }
     case CA_CMPLX64: {
       result = variance_ki_native_cmplx64(self, src, slab_axes, naxes, keep_axis, min_count);
+      RB_GC_GUARD(self);
       break;
     }
     case CA_CMPLX128: {
       result = variance_ki_native_cmplx128(self, src, slab_axes, naxes, keep_axis, min_count);
+      RB_GC_GUARD(self);
       break;
     }
     case CA_BOOLEAN: {
       result = variance_ki_native_bool(self, src, slab_axes, naxes, keep_axis, min_count);
+      RB_GC_GUARD(self);
       break;
     }
     case CA_OBJECT: {
       result = variance_ki_native_object(self, src, slab_axes, naxes, keep_axis, min_count);
+      RB_GC_GUARD(self);
       break;
     }
     default:
-      rb_raise(rb_eCADataTypeError, "variance_ki: source data_type :%s not supported (expected one of: i8, u8, i16, u16, i32, u32, i64, u64, f32, f64, cmplx64, cmplx128, bool, object)", ca_type_name[src->data_type]);
+      ca_face_refuse_numeric(self, ca_calling_method_name(), rb_eCADataTypeError);
+      rb_raise(rb_eCADataTypeError, "%s: source data_type :%s not supported (expected one of: int8, uint8, int16, uint16, int32, uint32, int64, uint64, float32, float64, cmplx64, cmplx128, boolean, object)", ca_calling_method_name(), ca_type_name[src->data_type]);
   }
   if ( ! NIL_P(rfval) ) {
-    if ( result == CA_UNDEF ) {
-      result = rfval;
-    } else if ( TYPE(result) != T_FLOAT && rb_obj_is_kind_of(result, rb_cCArray) ) {
-      CArray *cr;
-      GetCArray(result, cr);
-      if ( ca_has_mask(cr) ) {
-        result = rb_ca_mask_fill_copy(result, rfval);
-      }
-    }
+    result = ca_reduce_fill(result, rfval, fill_whole);
   }
   return result;
 }
@@ -1982,7 +1955,7 @@ stddevp_ki_native_i8 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, i
                                  slab_axes, naxes, 0);
   if ( rc != CA_ITER_OK ) {
     rb_raise(rb_eRuntimeError,
-             "stddevp_ki: kernel_iterator init failed rc=%d", rc);
+             "%s: kernel_iterator init failed rc=%d", ca_calling_method_name(), rc);
   }
   char       *p;
   boolean8_t *m;
@@ -1994,7 +1967,7 @@ stddevp_ki_native_i8 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, i
     ca_size_t masked_cnt = 0;
     CA_SLAB_REDUCE_SUM8_EX(int8_t, double, st, p, m, sum, 0.0, __stddevp_i8_P1EXPR, masked_cnt);
     ca_size_t n_valid = st.slab_elements - masked_cnt;
-    int __trigger = (min_count < 0 ? masked_cnt == st.slab_elements : n_valid < min_count);
+    int __trigger = ca_too_few_present(n_valid, min_count, 1);
     if ( __trigger ) {
       if ( ! op_mask ) {
         ca_create_mask(co);
@@ -2002,18 +1975,13 @@ stddevp_ki_native_i8 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, i
       }
       op_mask[out_i] = 1;
       op[out_i++] = (double) 0.0;
-    } else if ( n_valid < 1 ) {
-      /* Legacy degenerate case (e.g. sample variance with 1 unmasked
-         cell): return zero rather than UNDEF, matching pre-migration
-         behaviour of `cnt > 1 ? formula : 0`. */
-      op[out_i++] = (double) 0.0;
     } else {
       double mean = sum / (double) n_valid;
       /* Pass 2: M2 = Σ (x - mean)² */
       double M2 = 0.0;
       ca_size_t __unused_mc = 0;
       CA_SLAB_REDUCE_SUM8_EX(int8_t, double, st, p, m, M2, 0.0, __stddevp_i8_P2EXPR, __unused_mc);
-      op[out_i++] = (double)(sqrt(fmax(M2 / (double)(n_valid), 0.0)));
+      op[out_i++] = (double)(ca_sqrt_clamp(M2 / (double)(n_valid)));
     }
   }
   ca_iter_state_finish(&st);
@@ -2044,7 +2012,7 @@ stddevp_ki_native_u8 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, i
                                  slab_axes, naxes, 0);
   if ( rc != CA_ITER_OK ) {
     rb_raise(rb_eRuntimeError,
-             "stddevp_ki: kernel_iterator init failed rc=%d", rc);
+             "%s: kernel_iterator init failed rc=%d", ca_calling_method_name(), rc);
   }
   char       *p;
   boolean8_t *m;
@@ -2056,7 +2024,7 @@ stddevp_ki_native_u8 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, i
     ca_size_t masked_cnt = 0;
     CA_SLAB_REDUCE_SUM8_EX(uint8_t, double, st, p, m, sum, 0.0, __stddevp_u8_P1EXPR, masked_cnt);
     ca_size_t n_valid = st.slab_elements - masked_cnt;
-    int __trigger = (min_count < 0 ? masked_cnt == st.slab_elements : n_valid < min_count);
+    int __trigger = ca_too_few_present(n_valid, min_count, 1);
     if ( __trigger ) {
       if ( ! op_mask ) {
         ca_create_mask(co);
@@ -2064,18 +2032,13 @@ stddevp_ki_native_u8 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, i
       }
       op_mask[out_i] = 1;
       op[out_i++] = (double) 0.0;
-    } else if ( n_valid < 1 ) {
-      /* Legacy degenerate case (e.g. sample variance with 1 unmasked
-         cell): return zero rather than UNDEF, matching pre-migration
-         behaviour of `cnt > 1 ? formula : 0`. */
-      op[out_i++] = (double) 0.0;
     } else {
       double mean = sum / (double) n_valid;
       /* Pass 2: M2 = Σ (x - mean)² */
       double M2 = 0.0;
       ca_size_t __unused_mc = 0;
       CA_SLAB_REDUCE_SUM8_EX(uint8_t, double, st, p, m, M2, 0.0, __stddevp_u8_P2EXPR, __unused_mc);
-      op[out_i++] = (double)(sqrt(fmax(M2 / (double)(n_valid), 0.0)));
+      op[out_i++] = (double)(ca_sqrt_clamp(M2 / (double)(n_valid)));
     }
   }
   ca_iter_state_finish(&st);
@@ -2106,7 +2069,7 @@ stddevp_ki_native_i16 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes,
                                  slab_axes, naxes, 0);
   if ( rc != CA_ITER_OK ) {
     rb_raise(rb_eRuntimeError,
-             "stddevp_ki: kernel_iterator init failed rc=%d", rc);
+             "%s: kernel_iterator init failed rc=%d", ca_calling_method_name(), rc);
   }
   char       *p;
   boolean8_t *m;
@@ -2118,7 +2081,7 @@ stddevp_ki_native_i16 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes,
     ca_size_t masked_cnt = 0;
     CA_SLAB_REDUCE_SUM8_EX(int16_t, double, st, p, m, sum, 0.0, __stddevp_i16_P1EXPR, masked_cnt);
     ca_size_t n_valid = st.slab_elements - masked_cnt;
-    int __trigger = (min_count < 0 ? masked_cnt == st.slab_elements : n_valid < min_count);
+    int __trigger = ca_too_few_present(n_valid, min_count, 1);
     if ( __trigger ) {
       if ( ! op_mask ) {
         ca_create_mask(co);
@@ -2126,18 +2089,13 @@ stddevp_ki_native_i16 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes,
       }
       op_mask[out_i] = 1;
       op[out_i++] = (double) 0.0;
-    } else if ( n_valid < 1 ) {
-      /* Legacy degenerate case (e.g. sample variance with 1 unmasked
-         cell): return zero rather than UNDEF, matching pre-migration
-         behaviour of `cnt > 1 ? formula : 0`. */
-      op[out_i++] = (double) 0.0;
     } else {
       double mean = sum / (double) n_valid;
       /* Pass 2: M2 = Σ (x - mean)² */
       double M2 = 0.0;
       ca_size_t __unused_mc = 0;
       CA_SLAB_REDUCE_SUM8_EX(int16_t, double, st, p, m, M2, 0.0, __stddevp_i16_P2EXPR, __unused_mc);
-      op[out_i++] = (double)(sqrt(fmax(M2 / (double)(n_valid), 0.0)));
+      op[out_i++] = (double)(ca_sqrt_clamp(M2 / (double)(n_valid)));
     }
   }
   ca_iter_state_finish(&st);
@@ -2168,7 +2126,7 @@ stddevp_ki_native_u16 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes,
                                  slab_axes, naxes, 0);
   if ( rc != CA_ITER_OK ) {
     rb_raise(rb_eRuntimeError,
-             "stddevp_ki: kernel_iterator init failed rc=%d", rc);
+             "%s: kernel_iterator init failed rc=%d", ca_calling_method_name(), rc);
   }
   char       *p;
   boolean8_t *m;
@@ -2180,7 +2138,7 @@ stddevp_ki_native_u16 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes,
     ca_size_t masked_cnt = 0;
     CA_SLAB_REDUCE_SUM8_EX(uint16_t, double, st, p, m, sum, 0.0, __stddevp_u16_P1EXPR, masked_cnt);
     ca_size_t n_valid = st.slab_elements - masked_cnt;
-    int __trigger = (min_count < 0 ? masked_cnt == st.slab_elements : n_valid < min_count);
+    int __trigger = ca_too_few_present(n_valid, min_count, 1);
     if ( __trigger ) {
       if ( ! op_mask ) {
         ca_create_mask(co);
@@ -2188,18 +2146,13 @@ stddevp_ki_native_u16 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes,
       }
       op_mask[out_i] = 1;
       op[out_i++] = (double) 0.0;
-    } else if ( n_valid < 1 ) {
-      /* Legacy degenerate case (e.g. sample variance with 1 unmasked
-         cell): return zero rather than UNDEF, matching pre-migration
-         behaviour of `cnt > 1 ? formula : 0`. */
-      op[out_i++] = (double) 0.0;
     } else {
       double mean = sum / (double) n_valid;
       /* Pass 2: M2 = Σ (x - mean)² */
       double M2 = 0.0;
       ca_size_t __unused_mc = 0;
       CA_SLAB_REDUCE_SUM8_EX(uint16_t, double, st, p, m, M2, 0.0, __stddevp_u16_P2EXPR, __unused_mc);
-      op[out_i++] = (double)(sqrt(fmax(M2 / (double)(n_valid), 0.0)));
+      op[out_i++] = (double)(ca_sqrt_clamp(M2 / (double)(n_valid)));
     }
   }
   ca_iter_state_finish(&st);
@@ -2230,7 +2183,7 @@ stddevp_ki_native_i32 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes,
                                  slab_axes, naxes, 0);
   if ( rc != CA_ITER_OK ) {
     rb_raise(rb_eRuntimeError,
-             "stddevp_ki: kernel_iterator init failed rc=%d", rc);
+             "%s: kernel_iterator init failed rc=%d", ca_calling_method_name(), rc);
   }
   char       *p;
   boolean8_t *m;
@@ -2242,7 +2195,7 @@ stddevp_ki_native_i32 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes,
     ca_size_t masked_cnt = 0;
     CA_SLAB_REDUCE_SUM8_EX(int32_t, double, st, p, m, sum, 0.0, __stddevp_i32_P1EXPR, masked_cnt);
     ca_size_t n_valid = st.slab_elements - masked_cnt;
-    int __trigger = (min_count < 0 ? masked_cnt == st.slab_elements : n_valid < min_count);
+    int __trigger = ca_too_few_present(n_valid, min_count, 1);
     if ( __trigger ) {
       if ( ! op_mask ) {
         ca_create_mask(co);
@@ -2250,18 +2203,13 @@ stddevp_ki_native_i32 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes,
       }
       op_mask[out_i] = 1;
       op[out_i++] = (double) 0.0;
-    } else if ( n_valid < 1 ) {
-      /* Legacy degenerate case (e.g. sample variance with 1 unmasked
-         cell): return zero rather than UNDEF, matching pre-migration
-         behaviour of `cnt > 1 ? formula : 0`. */
-      op[out_i++] = (double) 0.0;
     } else {
       double mean = sum / (double) n_valid;
       /* Pass 2: M2 = Σ (x - mean)² */
       double M2 = 0.0;
       ca_size_t __unused_mc = 0;
       CA_SLAB_REDUCE_SUM8_EX(int32_t, double, st, p, m, M2, 0.0, __stddevp_i32_P2EXPR, __unused_mc);
-      op[out_i++] = (double)(sqrt(fmax(M2 / (double)(n_valid), 0.0)));
+      op[out_i++] = (double)(ca_sqrt_clamp(M2 / (double)(n_valid)));
     }
   }
   ca_iter_state_finish(&st);
@@ -2292,7 +2240,7 @@ stddevp_ki_native_u32 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes,
                                  slab_axes, naxes, 0);
   if ( rc != CA_ITER_OK ) {
     rb_raise(rb_eRuntimeError,
-             "stddevp_ki: kernel_iterator init failed rc=%d", rc);
+             "%s: kernel_iterator init failed rc=%d", ca_calling_method_name(), rc);
   }
   char       *p;
   boolean8_t *m;
@@ -2304,7 +2252,7 @@ stddevp_ki_native_u32 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes,
     ca_size_t masked_cnt = 0;
     CA_SLAB_REDUCE_SUM8_EX(uint32_t, double, st, p, m, sum, 0.0, __stddevp_u32_P1EXPR, masked_cnt);
     ca_size_t n_valid = st.slab_elements - masked_cnt;
-    int __trigger = (min_count < 0 ? masked_cnt == st.slab_elements : n_valid < min_count);
+    int __trigger = ca_too_few_present(n_valid, min_count, 1);
     if ( __trigger ) {
       if ( ! op_mask ) {
         ca_create_mask(co);
@@ -2312,18 +2260,13 @@ stddevp_ki_native_u32 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes,
       }
       op_mask[out_i] = 1;
       op[out_i++] = (double) 0.0;
-    } else if ( n_valid < 1 ) {
-      /* Legacy degenerate case (e.g. sample variance with 1 unmasked
-         cell): return zero rather than UNDEF, matching pre-migration
-         behaviour of `cnt > 1 ? formula : 0`. */
-      op[out_i++] = (double) 0.0;
     } else {
       double mean = sum / (double) n_valid;
       /* Pass 2: M2 = Σ (x - mean)² */
       double M2 = 0.0;
       ca_size_t __unused_mc = 0;
       CA_SLAB_REDUCE_SUM8_EX(uint32_t, double, st, p, m, M2, 0.0, __stddevp_u32_P2EXPR, __unused_mc);
-      op[out_i++] = (double)(sqrt(fmax(M2 / (double)(n_valid), 0.0)));
+      op[out_i++] = (double)(ca_sqrt_clamp(M2 / (double)(n_valid)));
     }
   }
   ca_iter_state_finish(&st);
@@ -2354,7 +2297,7 @@ stddevp_ki_native_i64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes,
                                  slab_axes, naxes, 0);
   if ( rc != CA_ITER_OK ) {
     rb_raise(rb_eRuntimeError,
-             "stddevp_ki: kernel_iterator init failed rc=%d", rc);
+             "%s: kernel_iterator init failed rc=%d", ca_calling_method_name(), rc);
   }
   char       *p;
   boolean8_t *m;
@@ -2366,7 +2309,7 @@ stddevp_ki_native_i64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes,
     ca_size_t masked_cnt = 0;
     CA_SLAB_REDUCE_SUM8_EX(int64_t, double, st, p, m, sum, 0.0, __stddevp_i64_P1EXPR, masked_cnt);
     ca_size_t n_valid = st.slab_elements - masked_cnt;
-    int __trigger = (min_count < 0 ? masked_cnt == st.slab_elements : n_valid < min_count);
+    int __trigger = ca_too_few_present(n_valid, min_count, 1);
     if ( __trigger ) {
       if ( ! op_mask ) {
         ca_create_mask(co);
@@ -2374,18 +2317,13 @@ stddevp_ki_native_i64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes,
       }
       op_mask[out_i] = 1;
       op[out_i++] = (double) 0.0;
-    } else if ( n_valid < 1 ) {
-      /* Legacy degenerate case (e.g. sample variance with 1 unmasked
-         cell): return zero rather than UNDEF, matching pre-migration
-         behaviour of `cnt > 1 ? formula : 0`. */
-      op[out_i++] = (double) 0.0;
     } else {
       double mean = sum / (double) n_valid;
       /* Pass 2: M2 = Σ (x - mean)² */
       double M2 = 0.0;
       ca_size_t __unused_mc = 0;
       CA_SLAB_REDUCE_SUM8_EX(int64_t, double, st, p, m, M2, 0.0, __stddevp_i64_P2EXPR, __unused_mc);
-      op[out_i++] = (double)(sqrt(fmax(M2 / (double)(n_valid), 0.0)));
+      op[out_i++] = (double)(ca_sqrt_clamp(M2 / (double)(n_valid)));
     }
   }
   ca_iter_state_finish(&st);
@@ -2416,7 +2354,7 @@ stddevp_ki_native_u64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes,
                                  slab_axes, naxes, 0);
   if ( rc != CA_ITER_OK ) {
     rb_raise(rb_eRuntimeError,
-             "stddevp_ki: kernel_iterator init failed rc=%d", rc);
+             "%s: kernel_iterator init failed rc=%d", ca_calling_method_name(), rc);
   }
   char       *p;
   boolean8_t *m;
@@ -2428,7 +2366,7 @@ stddevp_ki_native_u64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes,
     ca_size_t masked_cnt = 0;
     CA_SLAB_REDUCE_SUM8_EX(uint64_t, double, st, p, m, sum, 0.0, __stddevp_u64_P1EXPR, masked_cnt);
     ca_size_t n_valid = st.slab_elements - masked_cnt;
-    int __trigger = (min_count < 0 ? masked_cnt == st.slab_elements : n_valid < min_count);
+    int __trigger = ca_too_few_present(n_valid, min_count, 1);
     if ( __trigger ) {
       if ( ! op_mask ) {
         ca_create_mask(co);
@@ -2436,18 +2374,13 @@ stddevp_ki_native_u64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes,
       }
       op_mask[out_i] = 1;
       op[out_i++] = (double) 0.0;
-    } else if ( n_valid < 1 ) {
-      /* Legacy degenerate case (e.g. sample variance with 1 unmasked
-         cell): return zero rather than UNDEF, matching pre-migration
-         behaviour of `cnt > 1 ? formula : 0`. */
-      op[out_i++] = (double) 0.0;
     } else {
       double mean = sum / (double) n_valid;
       /* Pass 2: M2 = Σ (x - mean)² */
       double M2 = 0.0;
       ca_size_t __unused_mc = 0;
       CA_SLAB_REDUCE_SUM8_EX(uint64_t, double, st, p, m, M2, 0.0, __stddevp_u64_P2EXPR, __unused_mc);
-      op[out_i++] = (double)(sqrt(fmax(M2 / (double)(n_valid), 0.0)));
+      op[out_i++] = (double)(ca_sqrt_clamp(M2 / (double)(n_valid)));
     }
   }
   ca_iter_state_finish(&st);
@@ -2478,7 +2411,7 @@ stddevp_ki_native_f32 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes,
                                  slab_axes, naxes, 0);
   if ( rc != CA_ITER_OK ) {
     rb_raise(rb_eRuntimeError,
-             "stddevp_ki: kernel_iterator init failed rc=%d", rc);
+             "%s: kernel_iterator init failed rc=%d", ca_calling_method_name(), rc);
   }
   char       *p;
   boolean8_t *m;
@@ -2490,7 +2423,7 @@ stddevp_ki_native_f32 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes,
     ca_size_t masked_cnt = 0;
     CA_SLAB_REDUCE_SUM8_EX(float, double, st, p, m, sum, 0.0, __stddevp_f32_P1EXPR, masked_cnt);
     ca_size_t n_valid = st.slab_elements - masked_cnt;
-    int __trigger = (min_count < 0 ? masked_cnt == st.slab_elements : n_valid < min_count);
+    int __trigger = ca_too_few_present(n_valid, min_count, 1);
     if ( __trigger ) {
       if ( ! op_mask ) {
         ca_create_mask(co);
@@ -2498,18 +2431,13 @@ stddevp_ki_native_f32 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes,
       }
       op_mask[out_i] = 1;
       op[out_i++] = (double) 0.0;
-    } else if ( n_valid < 1 ) {
-      /* Legacy degenerate case (e.g. sample variance with 1 unmasked
-         cell): return zero rather than UNDEF, matching pre-migration
-         behaviour of `cnt > 1 ? formula : 0`. */
-      op[out_i++] = (double) 0.0;
     } else {
       double mean = sum / (double) n_valid;
       /* Pass 2: M2 = Σ (x - mean)² */
       double M2 = 0.0;
       ca_size_t __unused_mc = 0;
       CA_SLAB_REDUCE_SUM8_EX(float, double, st, p, m, M2, 0.0, __stddevp_f32_P2EXPR, __unused_mc);
-      op[out_i++] = (double)(sqrt(fmax(M2 / (double)(n_valid), 0.0)));
+      op[out_i++] = (double)(ca_sqrt_clamp(M2 / (double)(n_valid)));
     }
   }
   ca_iter_state_finish(&st);
@@ -2540,7 +2468,7 @@ stddevp_ki_native_f64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes,
                                  slab_axes, naxes, 0);
   if ( rc != CA_ITER_OK ) {
     rb_raise(rb_eRuntimeError,
-             "stddevp_ki: kernel_iterator init failed rc=%d", rc);
+             "%s: kernel_iterator init failed rc=%d", ca_calling_method_name(), rc);
   }
   char       *p;
   boolean8_t *m;
@@ -2552,7 +2480,7 @@ stddevp_ki_native_f64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes,
     ca_size_t masked_cnt = 0;
     CA_SLAB_REDUCE_SUM8_EX(double, double, st, p, m, sum, 0.0, __stddevp_f64_P1EXPR, masked_cnt);
     ca_size_t n_valid = st.slab_elements - masked_cnt;
-    int __trigger = (min_count < 0 ? masked_cnt == st.slab_elements : n_valid < min_count);
+    int __trigger = ca_too_few_present(n_valid, min_count, 1);
     if ( __trigger ) {
       if ( ! op_mask ) {
         ca_create_mask(co);
@@ -2560,18 +2488,13 @@ stddevp_ki_native_f64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes,
       }
       op_mask[out_i] = 1;
       op[out_i++] = (double) 0.0;
-    } else if ( n_valid < 1 ) {
-      /* Legacy degenerate case (e.g. sample variance with 1 unmasked
-         cell): return zero rather than UNDEF, matching pre-migration
-         behaviour of `cnt > 1 ? formula : 0`. */
-      op[out_i++] = (double) 0.0;
     } else {
       double mean = sum / (double) n_valid;
       /* Pass 2: M2 = Σ (x - mean)² */
       double M2 = 0.0;
       ca_size_t __unused_mc = 0;
       CA_SLAB_REDUCE_SUM8_EX(double, double, st, p, m, M2, 0.0, __stddevp_f64_P2EXPR, __unused_mc);
-      op[out_i++] = (double)(sqrt(fmax(M2 / (double)(n_valid), 0.0)));
+      op[out_i++] = (double)(ca_sqrt_clamp(M2 / (double)(n_valid)));
     }
   }
   ca_iter_state_finish(&st);
@@ -2602,7 +2525,7 @@ stddevp_ki_native_cmplx64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t nax
                                  slab_axes, naxes, 0);
   if ( rc != CA_ITER_OK ) {
     rb_raise(rb_eRuntimeError,
-             "stddevp_ki: kernel_iterator init failed rc=%d", rc);
+             "%s: kernel_iterator init failed rc=%d", ca_calling_method_name(), rc);
   }
   char       *p;
   boolean8_t *m;
@@ -2614,7 +2537,7 @@ stddevp_ki_native_cmplx64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t nax
     ca_size_t masked_cnt = 0;
     CA_SLAB_REDUCE_SUM8_EX(cmplx64_t, cmplx64_t, st, p, m, sum, 0, __stddevp_cmplx64_P1EXPR, masked_cnt);
     ca_size_t n_valid = st.slab_elements - masked_cnt;
-    int __trigger = (min_count < 0 ? masked_cnt == st.slab_elements : n_valid < min_count);
+    int __trigger = ca_too_few_present(n_valid, min_count, 1);
     if ( __trigger ) {
       if ( ! op_mask ) {
         ca_create_mask(co);
@@ -2622,18 +2545,13 @@ stddevp_ki_native_cmplx64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t nax
       }
       op_mask[out_i] = 1;
       op[out_i++] = (double) 0.0;
-    } else if ( n_valid < 1 ) {
-      /* Legacy degenerate case (e.g. sample variance with 1 unmasked
-         cell): return zero rather than UNDEF, matching pre-migration
-         behaviour of `cnt > 1 ? formula : 0`. */
-      op[out_i++] = (double) 0.0;
     } else {
       cmplx64_t mean = sum / (cmplx64_t)(double) n_valid;
       /* Pass 2: M2 = Σ (x - mean)² */
       double M2 = 0.0;
       ca_size_t __unused_mc = 0;
       CA_SLAB_REDUCE_SUM8_EX(cmplx64_t, double, st, p, m, M2, 0.0, __stddevp_cmplx64_P2EXPR, __unused_mc);
-      op[out_i++] = (double)(sqrt(fmax(M2 / (double)(n_valid), 0.0)));
+      op[out_i++] = (double)(ca_sqrt_clamp(M2 / (double)(n_valid)));
     }
   }
   ca_iter_state_finish(&st);
@@ -2664,7 +2582,7 @@ stddevp_ki_native_cmplx128 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t na
                                  slab_axes, naxes, 0);
   if ( rc != CA_ITER_OK ) {
     rb_raise(rb_eRuntimeError,
-             "stddevp_ki: kernel_iterator init failed rc=%d", rc);
+             "%s: kernel_iterator init failed rc=%d", ca_calling_method_name(), rc);
   }
   char       *p;
   boolean8_t *m;
@@ -2676,7 +2594,7 @@ stddevp_ki_native_cmplx128 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t na
     ca_size_t masked_cnt = 0;
     CA_SLAB_REDUCE_SUM8_EX(cmplx128_t, cmplx128_t, st, p, m, sum, 0, __stddevp_cmplx128_P1EXPR, masked_cnt);
     ca_size_t n_valid = st.slab_elements - masked_cnt;
-    int __trigger = (min_count < 0 ? masked_cnt == st.slab_elements : n_valid < min_count);
+    int __trigger = ca_too_few_present(n_valid, min_count, 1);
     if ( __trigger ) {
       if ( ! op_mask ) {
         ca_create_mask(co);
@@ -2684,18 +2602,13 @@ stddevp_ki_native_cmplx128 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t na
       }
       op_mask[out_i] = 1;
       op[out_i++] = (double) 0.0;
-    } else if ( n_valid < 1 ) {
-      /* Legacy degenerate case (e.g. sample variance with 1 unmasked
-         cell): return zero rather than UNDEF, matching pre-migration
-         behaviour of `cnt > 1 ? formula : 0`. */
-      op[out_i++] = (double) 0.0;
     } else {
       cmplx128_t mean = sum / (cmplx128_t)(double) n_valid;
       /* Pass 2: M2 = Σ (x - mean)² */
       double M2 = 0.0;
       ca_size_t __unused_mc = 0;
       CA_SLAB_REDUCE_SUM8_EX(cmplx128_t, double, st, p, m, M2, 0.0, __stddevp_cmplx128_P2EXPR, __unused_mc);
-      op[out_i++] = (double)(sqrt(fmax(M2 / (double)(n_valid), 0.0)));
+      op[out_i++] = (double)(ca_sqrt_clamp(M2 / (double)(n_valid)));
     }
   }
   ca_iter_state_finish(&st);
@@ -2726,7 +2639,7 @@ stddevp_ki_native_bool (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes,
                                  slab_axes, naxes, 0);
   if ( rc != CA_ITER_OK ) {
     rb_raise(rb_eRuntimeError,
-             "stddevp_ki: kernel_iterator init failed rc=%d", rc);
+             "%s: kernel_iterator init failed rc=%d", ca_calling_method_name(), rc);
   }
   char       *p;
   boolean8_t *m;
@@ -2738,7 +2651,7 @@ stddevp_ki_native_bool (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes,
     ca_size_t masked_cnt = 0;
     CA_SLAB_REDUCE_SUM8_EX(boolean8_t, double, st, p, m, sum, 0.0, __stddevp_bool_P1EXPR, masked_cnt);
     ca_size_t n_valid = st.slab_elements - masked_cnt;
-    int __trigger = (min_count < 0 ? masked_cnt == st.slab_elements : n_valid < min_count);
+    int __trigger = ca_too_few_present(n_valid, min_count, 1);
     if ( __trigger ) {
       if ( ! op_mask ) {
         ca_create_mask(co);
@@ -2746,18 +2659,13 @@ stddevp_ki_native_bool (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes,
       }
       op_mask[out_i] = 1;
       op[out_i++] = (double) 0.0;
-    } else if ( n_valid < 1 ) {
-      /* Legacy degenerate case (e.g. sample variance with 1 unmasked
-         cell): return zero rather than UNDEF, matching pre-migration
-         behaviour of `cnt > 1 ? formula : 0`. */
-      op[out_i++] = (double) 0.0;
     } else {
       double mean = sum / (double) n_valid;
       /* Pass 2: M2 = Σ (x - mean)² */
       double M2 = 0.0;
       ca_size_t __unused_mc = 0;
       CA_SLAB_REDUCE_SUM8_EX(boolean8_t, double, st, p, m, M2, 0.0, __stddevp_bool_P2EXPR, __unused_mc);
-      op[out_i++] = (double)(sqrt(fmax(M2 / (double)(n_valid), 0.0)));
+      op[out_i++] = (double)(ca_sqrt_clamp(M2 / (double)(n_valid)));
     }
   }
   ca_iter_state_finish(&st);
@@ -2774,19 +2682,45 @@ stddevp_ki_native_bool (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes,
 #undef __stddevp_bool_P1EXPR
 #undef __stddevp_bool_P2EXPR
+/* The object lane calls Ruby for every cell, so a walk can raise part
+   way.  Its states live in the caller's frame, and ca_iter_ensure
+   finishes them however the body leaves. */
+typedef struct {
+  ca_iter_state states[1];
+  VALUE self;
+  CArray *ca;
+  int8_t *slab_axes;
+  int8_t naxes;
+  int keep_axis;
+  ca_size_t min_count;
+} stddevp_ki_native_object_frame;
 static VALUE
-stddevp_ki_native_object (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int keep_axis, ca_size_t min_count)
+stddevp_ki_native_object_body (VALUE arg)
 {
+  stddevp_ki_native_object_frame *fr = (stddevp_ki_native_object_frame *) arg;
+  VALUE self = fr->self;
+  (void) self;
+  CArray *ca = fr->ca;
+  (void) ca;
+  int8_t *slab_axes = fr->slab_axes;
+  (void) slab_axes;
+  int8_t naxes = fr->naxes;
+  (void) naxes;
+  int keep_axis = fr->keep_axis;
+  (void) keep_axis;
+  ca_size_t min_count = fr->min_count;
+  (void) min_count;
+#define st (fr->states[0])
   VALUE   vout  = rb_ca_new_reduced(self, slab_axes, naxes, CA_FLOAT64, keep_axis);
   CArray *co;
   GetCArray(vout, co);
   double *op = (double *) co->ptr;
-  ca_iter_state st;
   int rc = ca_iter_state_init_l2(&st, ca, CA_SLAB_AXES,
                                  slab_axes, naxes, 0);
   if ( rc != CA_ITER_OK ) {
     rb_raise(rb_eRuntimeError,
-             "stddevp_ki: kernel_iterator init failed rc=%d", rc);
+             "%s: kernel_iterator init failed rc=%d", ca_calling_method_name(), rc);
   }
   char       *p;
   boolean8_t *m;
@@ -2798,7 +2732,7 @@ stddevp_ki_native_object (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxe
     ca_size_t masked_cnt = 0;
     CA_SLAB_REDUCE_T_EX(VALUE, st, p, m, sum, INT2FIX(0), sum = rb_funcall(sum, rb_intern("+"), 1, v), masked_cnt);
     ca_size_t n_valid = st.slab_elements - masked_cnt;
-    int __trigger = (min_count < 0 ? masked_cnt == st.slab_elements : n_valid < min_count);
+    int __trigger = ca_too_few_present(n_valid, min_count, 1);
     if ( __trigger ) {
       if ( ! op_mask ) {
         ca_create_mask(co);
@@ -2806,11 +2740,6 @@ stddevp_ki_native_object (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxe
       }
       op_mask[out_i] = 1;
       op[out_i++] = (double) 0.0;
-    } else if ( n_valid < 1 ) {
-      /* Legacy degenerate case (e.g. sample variance with 1 unmasked
-         cell): return zero rather than UNDEF, matching pre-migration
-         behaviour of `cnt > 1 ? formula : 0`. */
-      op[out_i++] = (double) 0.0;
     } else {
       VALUE mean = rb_funcall(sum, rb_intern("/"), 1, LONG2NUM(n_valid));
       /* Pass 2: M2 = Σ (x - mean)² */
@@ -2818,7 +2747,7 @@ stddevp_ki_native_object (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxe
       ca_size_t __unused_mc = 0;
       CA_SLAB_REDUCE_T_EX(VALUE, st, p, m, M2, INT2FIX(0), VALUE __d = rb_funcall(v, rb_intern("-"), 1, mean);
 M2 = rb_funcall(M2, rb_intern("+"), 1, rb_funcall(__d, rb_intern("*"), 1, __d)), __unused_mc);
-      op[out_i++] = (double)(sqrt(fmax(NUM2DBL(rb_funcall(M2, rb_intern("/"), 1, LONG2NUM(n_valid))), 0.0)));
+      op[out_i++] = (double)(ca_sqrt_clamp(NUM2DBL(rb_funcall(M2, rb_intern("/"), 1, LONG2NUM(n_valid)))));
     }
   }
   ca_iter_state_finish(&st);
@@ -2831,8 +2760,23 @@ M2 = rb_funcall(M2, rb_intern("+"), 1, rb_funcall(__d, rb_intern("*"), 1, __d)),
     return rb_float_new(op[0]);
   }
   return vout;
+#undef st
 }
+static VALUE
+stddevp_ki_native_object (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int keep_axis, ca_size_t min_count)
+{
+  stddevp_ki_native_object_frame fr;
+  fr.self = self;
+  fr.ca = ca;
+  fr.slab_axes = slab_axes;
+  fr.naxes = naxes;
+  fr.keep_axis = keep_axis;
+  fr.min_count = min_count;
+  return ca_iter_ensure(1, fr.states, stddevp_ki_native_object_body, (VALUE) &fr);
+}
 VALUE
 rb_ca_stddevp_ki (int argc, VALUE *argv, VALUE self)
 {
@@ -2844,89 +2788,101 @@ rb_ca_stddevp_ki (int argc, VALUE *argv, VALUE self)
   int keep_axis = RTEST(rkeep_axis);
   int8_t slab_axes[CA_RANK_MAX];
   int8_t naxes = rb_ca_parse_reduce_axes_kw(raxis, src, slab_axes);
-  ca_size_t min_count = -1;
-  if ( ! NIL_P(rmin_count) ) {
-    ca_size_t mc_user = NUM2SIZE(rmin_count);
-    if ( mc_user < 0 ) {
-      rb_raise(rb_eArgError, "stddevp_ki: :min_count must be non-negative, got %lld", (long long) mc_user);
+  if ( ca_is_lazy_view(src) ) {
+    volatile VALUE evaluated = ca_lazy_evaluated(self);
+    if ( ! NIL_P(evaluated) ) {
+      self = evaluated;
+      GetCArray(self, src);
     }
-    /* :min_count = 0 collapses to legacy default (any 1 valid). */
-    if ( mc_user > 0 ) min_count = mc_user;
+  }
+  ca_size_t min_count = ca_min_count_arg(rmin_count, NULL);
+  if ( rfval == CA_UNDEF ) rfval = Qnil;
+  int fill_whole = 0;
+  if ( ! NIL_P(rfval) && naxes == src->ndim && ! keep_axis ) {
+    keep_axis  = 1;
+    fill_whole = 1;
   }
   VALUE result;
   if ( argc > 0 ) {
-    rb_raise(rb_eArgError, "stddevp_ki: positional axis arguments are no longer accepted (got %d); use axis: kwarg, e.g. a.stddevp(axis: 0) or a.stddevp(axis: [0, 1])", argc);
+    rb_raise(rb_eArgError, "%s: positional axis arguments are no longer accepted (got %d); use the axis: keyword, e.g. a.%s(axis: 0) or a.%s(axis: [0, 1])", ca_calling_method_name(), argc, ca_calling_method_name(), ca_calling_method_name());
   }
   switch ( src->data_type ) {
     case CA_INT8: {
       result = stddevp_ki_native_i8(self, src, slab_axes, naxes, keep_axis, min_count);
+      RB_GC_GUARD(self);
       break;
     }
     case CA_UINT8: {
       result = stddevp_ki_native_u8(self, src, slab_axes, naxes, keep_axis, min_count);
+      RB_GC_GUARD(self);
       break;
     }
     case CA_INT16: {
       result = stddevp_ki_native_i16(self, src, slab_axes, naxes, keep_axis, min_count);
+      RB_GC_GUARD(self);
       break;
     }
     case CA_UINT16: {
       result = stddevp_ki_native_u16(self, src, slab_axes, naxes, keep_axis, min_count);
+      RB_GC_GUARD(self);
       break;
     }
     case CA_INT32: {
       result = stddevp_ki_native_i32(self, src, slab_axes, naxes, keep_axis, min_count);
+      RB_GC_GUARD(self);
       break;
     }
     case CA_UINT32: {
       result = stddevp_ki_native_u32(self, src, slab_axes, naxes, keep_axis, min_count);
+      RB_GC_GUARD(self);
       break;
     }
     case CA_INT64: {
       result = stddevp_ki_native_i64(self, src, slab_axes, naxes, keep_axis, min_count);
+      RB_GC_GUARD(self);
       break;
     }
     case CA_UINT64: {
       result = stddevp_ki_native_u64(self, src, slab_axes, naxes, keep_axis, min_count);
+      RB_GC_GUARD(self);
       break;
     }
     case CA_FLOAT32: {
       result = stddevp_ki_native_f32(self, src, slab_axes, naxes, keep_axis, min_count);
+      RB_GC_GUARD(self);
       break;
     }
     case CA_FLOAT64: {
       result = stddevp_ki_native_f64(self, src, slab_axes, naxes, keep_axis, min_count);
+      RB_GC_GUARD(self);
       break;
     }
     case CA_CMPLX64: {
       result = stddevp_ki_native_cmplx64(self, src, slab_axes, naxes, keep_axis, min_count);
+      RB_GC_GUARD(self);
       break;
     }
     case CA_CMPLX128: {
       result = stddevp_ki_native_cmplx128(self, src, slab_axes, naxes, keep_axis, min_count);
+      RB_GC_GUARD(self);
       break;
     }
     case CA_BOOLEAN: {
       result = stddevp_ki_native_bool(self, src, slab_axes, naxes, keep_axis, min_count);
+      RB_GC_GUARD(self);
       break;
     }
     case CA_OBJECT: {
       result = stddevp_ki_native_object(self, src, slab_axes, naxes, keep_axis, min_count);
+      RB_GC_GUARD(self);
       break;
     }
     default:
-      rb_raise(rb_eCADataTypeError, "stddevp_ki: source data_type :%s not supported (expected one of: i8, u8, i16, u16, i32, u32, i64, u64, f32, f64, cmplx64, cmplx128, bool, object)", ca_type_name[src->data_type]);
+      ca_face_refuse_numeric(self, ca_calling_method_name(), rb_eCADataTypeError);
+      rb_raise(rb_eCADataTypeError, "%s: source data_type :%s not supported (expected one of: int8, uint8, int16, uint16, int32, uint32, int64, uint64, float32, float64, cmplx64, cmplx128, boolean, object)", ca_calling_method_name(), ca_type_name[src->data_type]);
   }
   if ( ! NIL_P(rfval) ) {
-    if ( result == CA_UNDEF ) {
-      result = rfval;
-    } else if ( TYPE(result) != T_FLOAT && rb_obj_is_kind_of(result, rb_cCArray) ) {
-      CArray *cr;
-      GetCArray(result, cr);
-      if ( ca_has_mask(cr) ) {
-        result = rb_ca_mask_fill_copy(result, rfval);
-      }
-    }
+    result = ca_reduce_fill(result, rfval, fill_whole);
   }
   return result;
 }
@@ -2947,7 +2903,7 @@ stddev_ki_native_i8 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, in
                                  slab_axes, naxes, 0);
   if ( rc != CA_ITER_OK ) {
     rb_raise(rb_eRuntimeError,
-             "stddev_ki: kernel_iterator init failed rc=%d", rc);
+             "%s: kernel_iterator init failed rc=%d", ca_calling_method_name(), rc);
   }
   char       *p;
   boolean8_t *m;
@@ -2959,7 +2915,7 @@ stddev_ki_native_i8 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, in
     ca_size_t masked_cnt = 0;
     CA_SLAB_REDUCE_SUM8_EX(int8_t, double, st, p, m, sum, 0.0, __stddev_i8_P1EXPR, masked_cnt);
     ca_size_t n_valid = st.slab_elements - masked_cnt;
-    int __trigger = (min_count < 0 ? masked_cnt == st.slab_elements : n_valid < min_count);
+    int __trigger = ca_too_few_present(n_valid, min_count, 2);
     if ( __trigger ) {
       if ( ! op_mask ) {
         ca_create_mask(co);
@@ -2967,18 +2923,13 @@ stddev_ki_native_i8 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, in
       }
       op_mask[out_i] = 1;
       op[out_i++] = (double) 0.0;
-    } else if ( n_valid < 2 ) {
-      /* Legacy degenerate case (e.g. sample variance with 1 unmasked
-         cell): return zero rather than UNDEF, matching pre-migration
-         behaviour of `cnt > 1 ? formula : 0`. */
-      op[out_i++] = (double) 0.0;
     } else {
       double mean = sum / (double) n_valid;
       /* Pass 2: M2 = Σ (x - mean)² */
       double M2 = 0.0;
       ca_size_t __unused_mc = 0;
       CA_SLAB_REDUCE_SUM8_EX(int8_t, double, st, p, m, M2, 0.0, __stddev_i8_P2EXPR, __unused_mc);
-      op[out_i++] = (double)(sqrt(fmax(M2 / (double)(n_valid - 1), 0.0)));
+      op[out_i++] = (double)(ca_sqrt_clamp(M2 / (double)(n_valid - 1)));
     }
   }
   ca_iter_state_finish(&st);
@@ -3009,7 +2960,7 @@ stddev_ki_native_u8 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, in
                                  slab_axes, naxes, 0);
   if ( rc != CA_ITER_OK ) {
     rb_raise(rb_eRuntimeError,
-             "stddev_ki: kernel_iterator init failed rc=%d", rc);
+             "%s: kernel_iterator init failed rc=%d", ca_calling_method_name(), rc);
   }
   char       *p;
   boolean8_t *m;
@@ -3021,7 +2972,7 @@ stddev_ki_native_u8 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, in
     ca_size_t masked_cnt = 0;
     CA_SLAB_REDUCE_SUM8_EX(uint8_t, double, st, p, m, sum, 0.0, __stddev_u8_P1EXPR, masked_cnt);
     ca_size_t n_valid = st.slab_elements - masked_cnt;
-    int __trigger = (min_count < 0 ? masked_cnt == st.slab_elements : n_valid < min_count);
+    int __trigger = ca_too_few_present(n_valid, min_count, 2);
     if ( __trigger ) {
       if ( ! op_mask ) {
         ca_create_mask(co);
@@ -3029,18 +2980,13 @@ stddev_ki_native_u8 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, in
       }
       op_mask[out_i] = 1;
       op[out_i++] = (double) 0.0;
-    } else if ( n_valid < 2 ) {
-      /* Legacy degenerate case (e.g. sample variance with 1 unmasked
-         cell): return zero rather than UNDEF, matching pre-migration
-         behaviour of `cnt > 1 ? formula : 0`. */
-      op[out_i++] = (double) 0.0;
     } else {
       double mean = sum / (double) n_valid;
       /* Pass 2: M2 = Σ (x - mean)² */
       double M2 = 0.0;
       ca_size_t __unused_mc = 0;
       CA_SLAB_REDUCE_SUM8_EX(uint8_t, double, st, p, m, M2, 0.0, __stddev_u8_P2EXPR, __unused_mc);
-      op[out_i++] = (double)(sqrt(fmax(M2 / (double)(n_valid - 1), 0.0)));
+      op[out_i++] = (double)(ca_sqrt_clamp(M2 / (double)(n_valid - 1)));
     }
   }
   ca_iter_state_finish(&st);
@@ -3071,7 +3017,7 @@ stddev_ki_native_i16 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, i
                                  slab_axes, naxes, 0);
   if ( rc != CA_ITER_OK ) {
     rb_raise(rb_eRuntimeError,
-             "stddev_ki: kernel_iterator init failed rc=%d", rc);
+             "%s: kernel_iterator init failed rc=%d", ca_calling_method_name(), rc);
   }
   char       *p;
   boolean8_t *m;
@@ -3083,7 +3029,7 @@ stddev_ki_native_i16 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, i
     ca_size_t masked_cnt = 0;
     CA_SLAB_REDUCE_SUM8_EX(int16_t, double, st, p, m, sum, 0.0, __stddev_i16_P1EXPR, masked_cnt);
     ca_size_t n_valid = st.slab_elements - masked_cnt;
-    int __trigger = (min_count < 0 ? masked_cnt == st.slab_elements : n_valid < min_count);
+    int __trigger = ca_too_few_present(n_valid, min_count, 2);
     if ( __trigger ) {
       if ( ! op_mask ) {
         ca_create_mask(co);
@@ -3091,18 +3037,13 @@ stddev_ki_native_i16 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, i
       }
       op_mask[out_i] = 1;
       op[out_i++] = (double) 0.0;
-    } else if ( n_valid < 2 ) {
-      /* Legacy degenerate case (e.g. sample variance with 1 unmasked
-         cell): return zero rather than UNDEF, matching pre-migration
-         behaviour of `cnt > 1 ? formula : 0`. */
-      op[out_i++] = (double) 0.0;
     } else {
       double mean = sum / (double) n_valid;
       /* Pass 2: M2 = Σ (x - mean)² */
       double M2 = 0.0;
       ca_size_t __unused_mc = 0;
       CA_SLAB_REDUCE_SUM8_EX(int16_t, double, st, p, m, M2, 0.0, __stddev_i16_P2EXPR, __unused_mc);
-      op[out_i++] = (double)(sqrt(fmax(M2 / (double)(n_valid - 1), 0.0)));
+      op[out_i++] = (double)(ca_sqrt_clamp(M2 / (double)(n_valid - 1)));
     }
   }
   ca_iter_state_finish(&st);
@@ -3133,7 +3074,7 @@ stddev_ki_native_u16 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, i
                                  slab_axes, naxes, 0);
   if ( rc != CA_ITER_OK ) {
     rb_raise(rb_eRuntimeError,
-             "stddev_ki: kernel_iterator init failed rc=%d", rc);
+             "%s: kernel_iterator init failed rc=%d", ca_calling_method_name(), rc);
   }
   char       *p;
   boolean8_t *m;
@@ -3145,7 +3086,7 @@ stddev_ki_native_u16 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, i
     ca_size_t masked_cnt = 0;
     CA_SLAB_REDUCE_SUM8_EX(uint16_t, double, st, p, m, sum, 0.0, __stddev_u16_P1EXPR, masked_cnt);
     ca_size_t n_valid = st.slab_elements - masked_cnt;
-    int __trigger = (min_count < 0 ? masked_cnt == st.slab_elements : n_valid < min_count);
+    int __trigger = ca_too_few_present(n_valid, min_count, 2);
     if ( __trigger ) {
       if ( ! op_mask ) {
         ca_create_mask(co);
@@ -3153,18 +3094,13 @@ stddev_ki_native_u16 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, i
       }
       op_mask[out_i] = 1;
       op[out_i++] = (double) 0.0;
-    } else if ( n_valid < 2 ) {
-      /* Legacy degenerate case (e.g. sample variance with 1 unmasked
-         cell): return zero rather than UNDEF, matching pre-migration
-         behaviour of `cnt > 1 ? formula : 0`. */
-      op[out_i++] = (double) 0.0;
     } else {
       double mean = sum / (double) n_valid;
       /* Pass 2: M2 = Σ (x - mean)² */
       double M2 = 0.0;
       ca_size_t __unused_mc = 0;
       CA_SLAB_REDUCE_SUM8_EX(uint16_t, double, st, p, m, M2, 0.0, __stddev_u16_P2EXPR, __unused_mc);
-      op[out_i++] = (double)(sqrt(fmax(M2 / (double)(n_valid - 1), 0.0)));
+      op[out_i++] = (double)(ca_sqrt_clamp(M2 / (double)(n_valid - 1)));
     }
   }
   ca_iter_state_finish(&st);
@@ -3195,7 +3131,7 @@ stddev_ki_native_i32 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, i
                                  slab_axes, naxes, 0);
   if ( rc != CA_ITER_OK ) {
     rb_raise(rb_eRuntimeError,
-             "stddev_ki: kernel_iterator init failed rc=%d", rc);
+             "%s: kernel_iterator init failed rc=%d", ca_calling_method_name(), rc);
   }
   char       *p;
   boolean8_t *m;
@@ -3207,7 +3143,7 @@ stddev_ki_native_i32 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, i
     ca_size_t masked_cnt = 0;
     CA_SLAB_REDUCE_SUM8_EX(int32_t, double, st, p, m, sum, 0.0, __stddev_i32_P1EXPR, masked_cnt);
     ca_size_t n_valid = st.slab_elements - masked_cnt;
-    int __trigger = (min_count < 0 ? masked_cnt == st.slab_elements : n_valid < min_count);
+    int __trigger = ca_too_few_present(n_valid, min_count, 2);
     if ( __trigger ) {
       if ( ! op_mask ) {
         ca_create_mask(co);
@@ -3215,18 +3151,13 @@ stddev_ki_native_i32 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, i
       }
       op_mask[out_i] = 1;
       op[out_i++] = (double) 0.0;
-    } else if ( n_valid < 2 ) {
-      /* Legacy degenerate case (e.g. sample variance with 1 unmasked
-         cell): return zero rather than UNDEF, matching pre-migration
-         behaviour of `cnt > 1 ? formula : 0`. */
-      op[out_i++] = (double) 0.0;
     } else {
       double mean = sum / (double) n_valid;
       /* Pass 2: M2 = Σ (x - mean)² */
       double M2 = 0.0;
       ca_size_t __unused_mc = 0;
       CA_SLAB_REDUCE_SUM8_EX(int32_t, double, st, p, m, M2, 0.0, __stddev_i32_P2EXPR, __unused_mc);
-      op[out_i++] = (double)(sqrt(fmax(M2 / (double)(n_valid - 1), 0.0)));
+      op[out_i++] = (double)(ca_sqrt_clamp(M2 / (double)(n_valid - 1)));
     }
   }
   ca_iter_state_finish(&st);
@@ -3257,7 +3188,7 @@ stddev_ki_native_u32 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, i
                                  slab_axes, naxes, 0);
   if ( rc != CA_ITER_OK ) {
     rb_raise(rb_eRuntimeError,
-             "stddev_ki: kernel_iterator init failed rc=%d", rc);
+             "%s: kernel_iterator init failed rc=%d", ca_calling_method_name(), rc);
   }
   char       *p;
   boolean8_t *m;
@@ -3269,7 +3200,7 @@ stddev_ki_native_u32 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, i
     ca_size_t masked_cnt = 0;
     CA_SLAB_REDUCE_SUM8_EX(uint32_t, double, st, p, m, sum, 0.0, __stddev_u32_P1EXPR, masked_cnt);
     ca_size_t n_valid = st.slab_elements - masked_cnt;
-    int __trigger = (min_count < 0 ? masked_cnt == st.slab_elements : n_valid < min_count);
+    int __trigger = ca_too_few_present(n_valid, min_count, 2);
     if ( __trigger ) {
       if ( ! op_mask ) {
         ca_create_mask(co);
@@ -3277,18 +3208,13 @@ stddev_ki_native_u32 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, i
       }
       op_mask[out_i] = 1;
       op[out_i++] = (double) 0.0;
-    } else if ( n_valid < 2 ) {
-      /* Legacy degenerate case (e.g. sample variance with 1 unmasked
-         cell): return zero rather than UNDEF, matching pre-migration
-         behaviour of `cnt > 1 ? formula : 0`. */
-      op[out_i++] = (double) 0.0;
     } else {
       double mean = sum / (double) n_valid;
       /* Pass 2: M2 = Σ (x - mean)² */
       double M2 = 0.0;
       ca_size_t __unused_mc = 0;
       CA_SLAB_REDUCE_SUM8_EX(uint32_t, double, st, p, m, M2, 0.0, __stddev_u32_P2EXPR, __unused_mc);
-      op[out_i++] = (double)(sqrt(fmax(M2 / (double)(n_valid - 1), 0.0)));
+      op[out_i++] = (double)(ca_sqrt_clamp(M2 / (double)(n_valid - 1)));
     }
   }
   ca_iter_state_finish(&st);
@@ -3319,7 +3245,7 @@ stddev_ki_native_i64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, i
                                  slab_axes, naxes, 0);
   if ( rc != CA_ITER_OK ) {
     rb_raise(rb_eRuntimeError,
-             "stddev_ki: kernel_iterator init failed rc=%d", rc);
+             "%s: kernel_iterator init failed rc=%d", ca_calling_method_name(), rc);
   }
   char       *p;
   boolean8_t *m;
@@ -3331,7 +3257,7 @@ stddev_ki_native_i64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, i
     ca_size_t masked_cnt = 0;
     CA_SLAB_REDUCE_SUM8_EX(int64_t, double, st, p, m, sum, 0.0, __stddev_i64_P1EXPR, masked_cnt);
     ca_size_t n_valid = st.slab_elements - masked_cnt;
-    int __trigger = (min_count < 0 ? masked_cnt == st.slab_elements : n_valid < min_count);
+    int __trigger = ca_too_few_present(n_valid, min_count, 2);
     if ( __trigger ) {
       if ( ! op_mask ) {
         ca_create_mask(co);
@@ -3339,18 +3265,13 @@ stddev_ki_native_i64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, i
       }
       op_mask[out_i] = 1;
       op[out_i++] = (double) 0.0;
-    } else if ( n_valid < 2 ) {
-      /* Legacy degenerate case (e.g. sample variance with 1 unmasked
-         cell): return zero rather than UNDEF, matching pre-migration
-         behaviour of `cnt > 1 ? formula : 0`. */
-      op[out_i++] = (double) 0.0;
     } else {
       double mean = sum / (double) n_valid;
       /* Pass 2: M2 = Σ (x - mean)² */
       double M2 = 0.0;
       ca_size_t __unused_mc = 0;
       CA_SLAB_REDUCE_SUM8_EX(int64_t, double, st, p, m, M2, 0.0, __stddev_i64_P2EXPR, __unused_mc);
-      op[out_i++] = (double)(sqrt(fmax(M2 / (double)(n_valid - 1), 0.0)));
+      op[out_i++] = (double)(ca_sqrt_clamp(M2 / (double)(n_valid - 1)));
     }
   }
   ca_iter_state_finish(&st);
@@ -3381,7 +3302,7 @@ stddev_ki_native_u64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, i
                                  slab_axes, naxes, 0);
   if ( rc != CA_ITER_OK ) {
     rb_raise(rb_eRuntimeError,
-             "stddev_ki: kernel_iterator init failed rc=%d", rc);
+             "%s: kernel_iterator init failed rc=%d", ca_calling_method_name(), rc);
   }
   char       *p;
   boolean8_t *m;
@@ -3393,7 +3314,7 @@ stddev_ki_native_u64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, i
     ca_size_t masked_cnt = 0;
     CA_SLAB_REDUCE_SUM8_EX(uint64_t, double, st, p, m, sum, 0.0, __stddev_u64_P1EXPR, masked_cnt);
     ca_size_t n_valid = st.slab_elements - masked_cnt;
-    int __trigger = (min_count < 0 ? masked_cnt == st.slab_elements : n_valid < min_count);
+    int __trigger = ca_too_few_present(n_valid, min_count, 2);
     if ( __trigger ) {
       if ( ! op_mask ) {
         ca_create_mask(co);
@@ -3401,18 +3322,13 @@ stddev_ki_native_u64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, i
       }
       op_mask[out_i] = 1;
       op[out_i++] = (double) 0.0;
-    } else if ( n_valid < 2 ) {
-      /* Legacy degenerate case (e.g. sample variance with 1 unmasked
-         cell): return zero rather than UNDEF, matching pre-migration
-         behaviour of `cnt > 1 ? formula : 0`. */
-      op[out_i++] = (double) 0.0;
     } else {
       double mean = sum / (double) n_valid;
       /* Pass 2: M2 = Σ (x - mean)² */
       double M2 = 0.0;
       ca_size_t __unused_mc = 0;
       CA_SLAB_REDUCE_SUM8_EX(uint64_t, double, st, p, m, M2, 0.0, __stddev_u64_P2EXPR, __unused_mc);
-      op[out_i++] = (double)(sqrt(fmax(M2 / (double)(n_valid - 1), 0.0)));
+      op[out_i++] = (double)(ca_sqrt_clamp(M2 / (double)(n_valid - 1)));
     }
   }
   ca_iter_state_finish(&st);
@@ -3443,7 +3359,7 @@ stddev_ki_native_f32 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, i
                                  slab_axes, naxes, 0);
   if ( rc != CA_ITER_OK ) {
     rb_raise(rb_eRuntimeError,
-             "stddev_ki: kernel_iterator init failed rc=%d", rc);
+             "%s: kernel_iterator init failed rc=%d", ca_calling_method_name(), rc);
   }
   char       *p;
   boolean8_t *m;
@@ -3455,7 +3371,7 @@ stddev_ki_native_f32 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, i
     ca_size_t masked_cnt = 0;
     CA_SLAB_REDUCE_SUM8_EX(float, double, st, p, m, sum, 0.0, __stddev_f32_P1EXPR, masked_cnt);
     ca_size_t n_valid = st.slab_elements - masked_cnt;
-    int __trigger = (min_count < 0 ? masked_cnt == st.slab_elements : n_valid < min_count);
+    int __trigger = ca_too_few_present(n_valid, min_count, 2);
     if ( __trigger ) {
       if ( ! op_mask ) {
         ca_create_mask(co);
@@ -3463,18 +3379,13 @@ stddev_ki_native_f32 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, i
       }
       op_mask[out_i] = 1;
       op[out_i++] = (double) 0.0;
-    } else if ( n_valid < 2 ) {
-      /* Legacy degenerate case (e.g. sample variance with 1 unmasked
-         cell): return zero rather than UNDEF, matching pre-migration
-         behaviour of `cnt > 1 ? formula : 0`. */
-      op[out_i++] = (double) 0.0;
     } else {
       double mean = sum / (double) n_valid;
       /* Pass 2: M2 = Σ (x - mean)² */
       double M2 = 0.0;
       ca_size_t __unused_mc = 0;
       CA_SLAB_REDUCE_SUM8_EX(float, double, st, p, m, M2, 0.0, __stddev_f32_P2EXPR, __unused_mc);
-      op[out_i++] = (double)(sqrt(fmax(M2 / (double)(n_valid - 1), 0.0)));
+      op[out_i++] = (double)(ca_sqrt_clamp(M2 / (double)(n_valid - 1)));
     }
   }
   ca_iter_state_finish(&st);
@@ -3505,7 +3416,7 @@ stddev_ki_native_f64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, i
                                  slab_axes, naxes, 0);
   if ( rc != CA_ITER_OK ) {
     rb_raise(rb_eRuntimeError,
-             "stddev_ki: kernel_iterator init failed rc=%d", rc);
+             "%s: kernel_iterator init failed rc=%d", ca_calling_method_name(), rc);
   }
   char       *p;
   boolean8_t *m;
@@ -3517,7 +3428,7 @@ stddev_ki_native_f64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, i
     ca_size_t masked_cnt = 0;
     CA_SLAB_REDUCE_SUM8_EX(double, double, st, p, m, sum, 0.0, __stddev_f64_P1EXPR, masked_cnt);
     ca_size_t n_valid = st.slab_elements - masked_cnt;
-    int __trigger = (min_count < 0 ? masked_cnt == st.slab_elements : n_valid < min_count);
+    int __trigger = ca_too_few_present(n_valid, min_count, 2);
     if ( __trigger ) {
       if ( ! op_mask ) {
         ca_create_mask(co);
@@ -3525,18 +3436,13 @@ stddev_ki_native_f64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, i
       }
       op_mask[out_i] = 1;
       op[out_i++] = (double) 0.0;
-    } else if ( n_valid < 2 ) {
-      /* Legacy degenerate case (e.g. sample variance with 1 unmasked
-         cell): return zero rather than UNDEF, matching pre-migration
-         behaviour of `cnt > 1 ? formula : 0`. */
-      op[out_i++] = (double) 0.0;
     } else {
       double mean = sum / (double) n_valid;
       /* Pass 2: M2 = Σ (x - mean)² */
       double M2 = 0.0;
       ca_size_t __unused_mc = 0;
       CA_SLAB_REDUCE_SUM8_EX(double, double, st, p, m, M2, 0.0, __stddev_f64_P2EXPR, __unused_mc);
-      op[out_i++] = (double)(sqrt(fmax(M2 / (double)(n_valid - 1), 0.0)));
+      op[out_i++] = (double)(ca_sqrt_clamp(M2 / (double)(n_valid - 1)));
     }
   }
   ca_iter_state_finish(&st);
@@ -3567,7 +3473,7 @@ stddev_ki_native_cmplx64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxe
                                  slab_axes, naxes, 0);
   if ( rc != CA_ITER_OK ) {
     rb_raise(rb_eRuntimeError,
-             "stddev_ki: kernel_iterator init failed rc=%d", rc);
+             "%s: kernel_iterator init failed rc=%d", ca_calling_method_name(), rc);
   }
   char       *p;
   boolean8_t *m;
@@ -3579,7 +3485,7 @@ stddev_ki_native_cmplx64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxe
     ca_size_t masked_cnt = 0;
     CA_SLAB_REDUCE_SUM8_EX(cmplx64_t, cmplx64_t, st, p, m, sum, 0, __stddev_cmplx64_P1EXPR, masked_cnt);
     ca_size_t n_valid = st.slab_elements - masked_cnt;
-    int __trigger = (min_count < 0 ? masked_cnt == st.slab_elements : n_valid < min_count);
+    int __trigger = ca_too_few_present(n_valid, min_count, 2);
     if ( __trigger ) {
       if ( ! op_mask ) {
         ca_create_mask(co);
@@ -3587,18 +3493,13 @@ stddev_ki_native_cmplx64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxe
       }
       op_mask[out_i] = 1;
       op[out_i++] = (double) 0.0;
-    } else if ( n_valid < 2 ) {
-      /* Legacy degenerate case (e.g. sample variance with 1 unmasked
-         cell): return zero rather than UNDEF, matching pre-migration
-         behaviour of `cnt > 1 ? formula : 0`. */
-      op[out_i++] = (double) 0.0;
     } else {
       cmplx64_t mean = sum / (cmplx64_t)(double) n_valid;
       /* Pass 2: M2 = Σ (x - mean)² */
       double M2 = 0.0;
       ca_size_t __unused_mc = 0;
       CA_SLAB_REDUCE_SUM8_EX(cmplx64_t, double, st, p, m, M2, 0.0, __stddev_cmplx64_P2EXPR, __unused_mc);
-      op[out_i++] = (double)(sqrt(fmax(M2 / (double)(n_valid - 1), 0.0)));
+      op[out_i++] = (double)(ca_sqrt_clamp(M2 / (double)(n_valid - 1)));
     }
   }
   ca_iter_state_finish(&st);
@@ -3629,7 +3530,7 @@ stddev_ki_native_cmplx128 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t nax
                                  slab_axes, naxes, 0);
   if ( rc != CA_ITER_OK ) {
     rb_raise(rb_eRuntimeError,
-             "stddev_ki: kernel_iterator init failed rc=%d", rc);
+             "%s: kernel_iterator init failed rc=%d", ca_calling_method_name(), rc);
   }
   char       *p;
   boolean8_t *m;
@@ -3641,7 +3542,7 @@ stddev_ki_native_cmplx128 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t nax
     ca_size_t masked_cnt = 0;
     CA_SLAB_REDUCE_SUM8_EX(cmplx128_t, cmplx128_t, st, p, m, sum, 0, __stddev_cmplx128_P1EXPR, masked_cnt);
     ca_size_t n_valid = st.slab_elements - masked_cnt;
-    int __trigger = (min_count < 0 ? masked_cnt == st.slab_elements : n_valid < min_count);
+    int __trigger = ca_too_few_present(n_valid, min_count, 2);
     if ( __trigger ) {
       if ( ! op_mask ) {
         ca_create_mask(co);
@@ -3649,18 +3550,13 @@ stddev_ki_native_cmplx128 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t nax
       }
       op_mask[out_i] = 1;
       op[out_i++] = (double) 0.0;
-    } else if ( n_valid < 2 ) {
-      /* Legacy degenerate case (e.g. sample variance with 1 unmasked
-         cell): return zero rather than UNDEF, matching pre-migration
-         behaviour of `cnt > 1 ? formula : 0`. */
-      op[out_i++] = (double) 0.0;
     } else {
       cmplx128_t mean = sum / (cmplx128_t)(double) n_valid;
       /* Pass 2: M2 = Σ (x - mean)² */
       double M2 = 0.0;
       ca_size_t __unused_mc = 0;
       CA_SLAB_REDUCE_SUM8_EX(cmplx128_t, double, st, p, m, M2, 0.0, __stddev_cmplx128_P2EXPR, __unused_mc);
-      op[out_i++] = (double)(sqrt(fmax(M2 / (double)(n_valid - 1), 0.0)));
+      op[out_i++] = (double)(ca_sqrt_clamp(M2 / (double)(n_valid - 1)));
     }
   }
   ca_iter_state_finish(&st);
@@ -3691,7 +3587,7 @@ stddev_ki_native_bool (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes,
                                  slab_axes, naxes, 0);
   if ( rc != CA_ITER_OK ) {
     rb_raise(rb_eRuntimeError,
-             "stddev_ki: kernel_iterator init failed rc=%d", rc);
+             "%s: kernel_iterator init failed rc=%d", ca_calling_method_name(), rc);
   }
   char       *p;
   boolean8_t *m;
@@ -3703,7 +3599,7 @@ stddev_ki_native_bool (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes,
     ca_size_t masked_cnt = 0;
     CA_SLAB_REDUCE_SUM8_EX(boolean8_t, double, st, p, m, sum, 0.0, __stddev_bool_P1EXPR, masked_cnt);
     ca_size_t n_valid = st.slab_elements - masked_cnt;
-    int __trigger = (min_count < 0 ? masked_cnt == st.slab_elements : n_valid < min_count);
+    int __trigger = ca_too_few_present(n_valid, min_count, 2);
     if ( __trigger ) {
       if ( ! op_mask ) {
         ca_create_mask(co);
@@ -3711,18 +3607,13 @@ stddev_ki_native_bool (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes,
       }
       op_mask[out_i] = 1;
       op[out_i++] = (double) 0.0;
-    } else if ( n_valid < 2 ) {
-      /* Legacy degenerate case (e.g. sample variance with 1 unmasked
-         cell): return zero rather than UNDEF, matching pre-migration
-         behaviour of `cnt > 1 ? formula : 0`. */
-      op[out_i++] = (double) 0.0;
     } else {
       double mean = sum / (double) n_valid;
       /* Pass 2: M2 = Σ (x - mean)² */
       double M2 = 0.0;
       ca_size_t __unused_mc = 0;
       CA_SLAB_REDUCE_SUM8_EX(boolean8_t, double, st, p, m, M2, 0.0, __stddev_bool_P2EXPR, __unused_mc);
-      op[out_i++] = (double)(sqrt(fmax(M2 / (double)(n_valid - 1), 0.0)));
+      op[out_i++] = (double)(ca_sqrt_clamp(M2 / (double)(n_valid - 1)));
     }
   }
   ca_iter_state_finish(&st);
@@ -3739,19 +3630,45 @@ stddev_ki_native_bool (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes,
 #undef __stddev_bool_P1EXPR
 #undef __stddev_bool_P2EXPR
+/* The object lane calls Ruby for every cell, so a walk can raise part
+   way.  Its states live in the caller's frame, and ca_iter_ensure
+   finishes them however the body leaves. */
+typedef struct {
+  ca_iter_state states[1];
+  VALUE self;
+  CArray *ca;
+  int8_t *slab_axes;
+  int8_t naxes;
+  int keep_axis;
+  ca_size_t min_count;
+} stddev_ki_native_object_frame;
 static VALUE
-stddev_ki_native_object (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int keep_axis, ca_size_t min_count)
+stddev_ki_native_object_body (VALUE arg)
 {
+  stddev_ki_native_object_frame *fr = (stddev_ki_native_object_frame *) arg;
+  VALUE self = fr->self;
+  (void) self;
+  CArray *ca = fr->ca;
+  (void) ca;
+  int8_t *slab_axes = fr->slab_axes;
+  (void) slab_axes;
+  int8_t naxes = fr->naxes;
+  (void) naxes;
+  int keep_axis = fr->keep_axis;
+  (void) keep_axis;
+  ca_size_t min_count = fr->min_count;
+  (void) min_count;
+#define st (fr->states[0])
   VALUE   vout  = rb_ca_new_reduced(self, slab_axes, naxes, CA_FLOAT64, keep_axis);
   CArray *co;
   GetCArray(vout, co);
   double *op = (double *) co->ptr;
-  ca_iter_state st;
   int rc = ca_iter_state_init_l2(&st, ca, CA_SLAB_AXES,
                                  slab_axes, naxes, 0);
   if ( rc != CA_ITER_OK ) {
     rb_raise(rb_eRuntimeError,
-             "stddev_ki: kernel_iterator init failed rc=%d", rc);
+             "%s: kernel_iterator init failed rc=%d", ca_calling_method_name(), rc);
   }
   char       *p;
   boolean8_t *m;
@@ -3763,7 +3680,7 @@ stddev_ki_native_object (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes
     ca_size_t masked_cnt = 0;
     CA_SLAB_REDUCE_T_EX(VALUE, st, p, m, sum, INT2FIX(0), sum = rb_funcall(sum, rb_intern("+"), 1, v), masked_cnt);
     ca_size_t n_valid = st.slab_elements - masked_cnt;
-    int __trigger = (min_count < 0 ? masked_cnt == st.slab_elements : n_valid < min_count);
+    int __trigger = ca_too_few_present(n_valid, min_count, 2);
     if ( __trigger ) {
       if ( ! op_mask ) {
         ca_create_mask(co);
@@ -3771,11 +3688,6 @@ stddev_ki_native_object (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes
       }
       op_mask[out_i] = 1;
       op[out_i++] = (double) 0.0;
-    } else if ( n_valid < 2 ) {
-      /* Legacy degenerate case (e.g. sample variance with 1 unmasked
-         cell): return zero rather than UNDEF, matching pre-migration
-         behaviour of `cnt > 1 ? formula : 0`. */
-      op[out_i++] = (double) 0.0;
     } else {
       VALUE mean = rb_funcall(sum, rb_intern("/"), 1, LONG2NUM(n_valid));
       /* Pass 2: M2 = Σ (x - mean)² */
@@ -3783,7 +3695,7 @@ stddev_ki_native_object (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes
       ca_size_t __unused_mc = 0;
       CA_SLAB_REDUCE_T_EX(VALUE, st, p, m, M2, INT2FIX(0), VALUE __d = rb_funcall(v, rb_intern("-"), 1, mean);
 M2 = rb_funcall(M2, rb_intern("+"), 1, rb_funcall(__d, rb_intern("*"), 1, __d)), __unused_mc);
-      op[out_i++] = (double)(sqrt(fmax(NUM2DBL(rb_funcall(M2, rb_intern("/"), 1, LONG2NUM(n_valid - 1))), 0.0)));
+      op[out_i++] = (double)(ca_sqrt_clamp(NUM2DBL(rb_funcall(M2, rb_intern("/"), 1, LONG2NUM(n_valid - 1)))));
     }
   }
   ca_iter_state_finish(&st);
@@ -3796,8 +3708,23 @@ M2 = rb_funcall(M2, rb_intern("+"), 1, rb_funcall(__d, rb_intern("*"), 1, __d)),
     return rb_float_new(op[0]);
   }
   return vout;
+#undef st
 }
+static VALUE
+stddev_ki_native_object (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int keep_axis, ca_size_t min_count)
+{
+  stddev_ki_native_object_frame fr;
+  fr.self = self;
+  fr.ca = ca;
+  fr.slab_axes = slab_axes;
+  fr.naxes = naxes;
+  fr.keep_axis = keep_axis;
+  fr.min_count = min_count;
+  return ca_iter_ensure(1, fr.states, stddev_ki_native_object_body, (VALUE) &fr);
+}
 VALUE
 rb_ca_stddev_ki (int argc, VALUE *argv, VALUE self)
 {
@@ -3809,89 +3736,101 @@ rb_ca_stddev_ki (int argc, VALUE *argv, VALUE self)
   int keep_axis = RTEST(rkeep_axis);
   int8_t slab_axes[CA_RANK_MAX];
   int8_t naxes = rb_ca_parse_reduce_axes_kw(raxis, src, slab_axes);
-  ca_size_t min_count = -1;
-  if ( ! NIL_P(rmin_count) ) {
-    ca_size_t mc_user = NUM2SIZE(rmin_count);
-    if ( mc_user < 0 ) {
-      rb_raise(rb_eArgError, "stddev_ki: :min_count must be non-negative, got %lld", (long long) mc_user);
+  if ( ca_is_lazy_view(src) ) {
+    volatile VALUE evaluated = ca_lazy_evaluated(self);
+    if ( ! NIL_P(evaluated) ) {
+      self = evaluated;
+      GetCArray(self, src);
     }
-    /* :min_count = 0 collapses to legacy default (any 1 valid). */
-    if ( mc_user > 0 ) min_count = mc_user;
+  }
+  ca_size_t min_count = ca_min_count_arg(rmin_count, NULL);
+  if ( rfval == CA_UNDEF ) rfval = Qnil;
+  int fill_whole = 0;
+  if ( ! NIL_P(rfval) && naxes == src->ndim && ! keep_axis ) {
+    keep_axis  = 1;
+    fill_whole = 1;
   }
   VALUE result;
   if ( argc > 0 ) {
-    rb_raise(rb_eArgError, "stddev_ki: positional axis arguments are no longer accepted (got %d); use axis: kwarg, e.g. a.stddev(axis: 0) or a.stddev(axis: [0, 1])", argc);
+    rb_raise(rb_eArgError, "%s: positional axis arguments are no longer accepted (got %d); use the axis: keyword, e.g. a.%s(axis: 0) or a.%s(axis: [0, 1])", ca_calling_method_name(), argc, ca_calling_method_name(), ca_calling_method_name());
   }
   switch ( src->data_type ) {
     case CA_INT8: {
       result = stddev_ki_native_i8(self, src, slab_axes, naxes, keep_axis, min_count);
+      RB_GC_GUARD(self);
       break;
     }
     case CA_UINT8: {
       result = stddev_ki_native_u8(self, src, slab_axes, naxes, keep_axis, min_count);
+      RB_GC_GUARD(self);
       break;
     }
     case CA_INT16: {
       result = stddev_ki_native_i16(self, src, slab_axes, naxes, keep_axis, min_count);
+      RB_GC_GUARD(self);
       break;
     }
     case CA_UINT16: {
       result = stddev_ki_native_u16(self, src, slab_axes, naxes, keep_axis, min_count);
+      RB_GC_GUARD(self);
       break;
     }
     case CA_INT32: {
       result = stddev_ki_native_i32(self, src, slab_axes, naxes, keep_axis, min_count);
+      RB_GC_GUARD(self);
       break;
     }
     case CA_UINT32: {
       result = stddev_ki_native_u32(self, src, slab_axes, naxes, keep_axis, min_count);
+      RB_GC_GUARD(self);
       break;
     }
     case CA_INT64: {
       result = stddev_ki_native_i64(self, src, slab_axes, naxes, keep_axis, min_count);
+      RB_GC_GUARD(self);
       break;
     }
     case CA_UINT64: {
       result = stddev_ki_native_u64(self, src, slab_axes, naxes, keep_axis, min_count);
+      RB_GC_GUARD(self);
       break;
     }
     case CA_FLOAT32: {
       result = stddev_ki_native_f32(self, src, slab_axes, naxes, keep_axis, min_count);
+      RB_GC_GUARD(self);
       break;
     }
     case CA_FLOAT64: {
       result = stddev_ki_native_f64(self, src, slab_axes, naxes, keep_axis, min_count);
+      RB_GC_GUARD(self);
       break;
     }
     case CA_CMPLX64: {
       result = stddev_ki_native_cmplx64(self, src, slab_axes, naxes, keep_axis, min_count);
+      RB_GC_GUARD(self);
       break;
     }
     case CA_CMPLX128: {
       result = stddev_ki_native_cmplx128(self, src, slab_axes, naxes, keep_axis, min_count);
+      RB_GC_GUARD(self);
       break;
     }
     case CA_BOOLEAN: {
       result = stddev_ki_native_bool(self, src, slab_axes, naxes, keep_axis, min_count);
+      RB_GC_GUARD(self);
       break;
     }
     case CA_OBJECT: {
       result = stddev_ki_native_object(self, src, slab_axes, naxes, keep_axis, min_count);
+      RB_GC_GUARD(self);
       break;
     }
     default:
-      rb_raise(rb_eCADataTypeError, "stddev_ki: source data_type :%s not supported (expected one of: i8, u8, i16, u16, i32, u32, i64, u64, f32, f64, cmplx64, cmplx128, bool, object)", ca_type_name[src->data_type]);
+      ca_face_refuse_numeric(self, ca_calling_method_name(), rb_eCADataTypeError);
+      rb_raise(rb_eCADataTypeError, "%s: source data_type :%s not supported (expected one of: int8, uint8, int16, uint16, int32, uint32, int64, uint64, float32, float64, cmplx64, cmplx128, boolean, object)", ca_calling_method_name(), ca_type_name[src->data_type]);
   }
   if ( ! NIL_P(rfval) ) {
-    if ( result == CA_UNDEF ) {
-      result = rfval;
-    } else if ( TYPE(result) != T_FLOAT && rb_obj_is_kind_of(result, rb_cCArray) ) {
-      CArray *cr;
-      GetCArray(result, cr);
-      if ( ca_has_mask(cr) ) {
-        result = rb_ca_mask_fill_copy(result, rfval);
-      }
-    }
+    result = ca_reduce_fill(result, rfval, fill_whole);
   }
   return result;
 }