carray 2.0.0 → 3.0.0

data/ext/ca_axis_group.c

@@ -0,0 +1,1208 @@
+/* ---------------------------------------------------------------------------
+
+  Axis-group compute kernels: the grouped reduction and the grouped segment
+  scan.
+
+  INTERNAL kernels behind `__...__`-named methods.  The user-facing surface
+  (CACategorical / AxisGroup type gate in `[]`, CAGroupIterator, the :group
+  reduce dispatch, GroupLabels) is wired in ca_group_iter.c (the `[]` gate and
+  the iterator) + `axis_group` (lib/carray/axis_group.rb).  Both kernels are
+  written in the general form: several group axes + rank-N categorical (an N-D
+  codes map) via a per-slab-element composite code, native dtype dispatch,
+  mask support.  They take pre-built code bundles, so they are independent of
+  how the classifier is constructed.
+
+  Mechanism, shared by both kernels:
+    - CA_FOR_EACH_SLAB pins the union of grouped source axes as the slab and
+      walks the band (= non-grouped) axes in the outer iter.
+    - For each slab element a composite group code is computed on the fly from
+      the per-bundle code tables (per-axis code, flat-collapsed; a rank-N
+      categorical = one bundle consuming more than one source axis).  The
+      composite code is never materialised — it is a per-element local.
+    - reduce: scatter out[composite * band + b] op= value, sort-free.  Peak
+      memory = output + per-group accumulators (O(K*band)), never O(N).
+    - scan: emit the running value per element into a source-shaped output
+      (see the section comment above `rb_ca_axis_group_scan`).
+
+  Group codes outside [0, k) at any bundle mark the slab element unassigned
+  (skipped, contributing to no group) — mirrors digitize OOB.
+
+  Float reductions carry the CArray ε-close contract: relative error bounded,
+  not bit-exact.
+
+--------------------------------------------------------------------------- */
+
+#include "carray.h"
+#include "ca_kernel_iterator.h"
+#include <math.h>
+
+/* op codes */
+enum {
+  GR_SUM = 0, GR_PROD, GR_MEAN, GR_MIN, GR_MAX,
+  GR_VARIANCE, GR_STDDEV, GR_COUNT, GR_ALL, GR_ANY,
+  GR_VARIANCEP, GR_STDDEVP, GR_MINADDR, GR_MAXADDR
+};
+
+static int
+group_op_code (VALUE vop)
+{
+  ID id = SYM2ID(vop);
+  if      ( id == rb_intern("sum") )      return GR_SUM;
+  else if ( id == rb_intern("prod") )     return GR_PROD;
+  else if ( id == rb_intern("mean") )     return GR_MEAN;
+  else if ( id == rb_intern("min") )      return GR_MIN;
+  else if ( id == rb_intern("max") )      return GR_MAX;
+  else if ( id == rb_intern("variance") ) return GR_VARIANCE;
+  else if ( id == rb_intern("stddev") )   return GR_STDDEV;
+  else if ( id == rb_intern("variancep") ) return GR_VARIANCEP;
+  else if ( id == rb_intern("stddevp") )   return GR_STDDEVP;
+  else if ( id == rb_intern("count") )    return GR_COUNT;
+  else if ( id == rb_intern("count_not_masked") ) return GR_COUNT;
+  else if ( id == rb_intern("min_addr") ) return GR_MINADDR;
+  else if ( id == rb_intern("max_addr") ) return GR_MAXADDR;
+  else if ( id == rb_intern("all") )      return GR_ALL;
+  else if ( id == rb_intern("any") )      return GR_ANY;
+  rb_raise(rb_eArgError, "axis_group_reduce: unsupported op :%s",
+           rb_id2name(id));
+}
+
+/* ---- per-slab-element composite-code walk, templated on the load type ----
+
+   Reads `*(T *)(p + data_off)` for each slab element, computes the composite
+   group code, applies the mask, and folds the value into whichever
+   accumulator buffers are non-NULL.  Accumulation is always in `double` (or
+   counts in ca_size_t), so the op-finalisation below is type-agnostic — only
+   the LOAD is monomorphised per dtype, keeping the inner loop autovectorisable
+   while avoiding a forced float64 materialise of the (large) source. */
+
+/* GROUP_WALK(T, ACCUM): one CA_FOR_EACH_SLAB pass.  For each slab element it
+   computes the composite group code, applies the mask, then runs ACCUM with
+   `v` (the element, widened to double) and `o` (the output flat index =
+   code * band + band_flat) in scope.  ACCUM is the only per-op-varying part,
+   so the dtype is monomorphised once per type while sum / mean / variance /
+   ... reuse the same walk. */
+/* Sentinel o-code for a slab element whose composite code is out of range
+   (excluded categorical), stored in the precomputed plan below. */
+#define GW_SKIP (~(ca_size_t) 0)
+
+/* CAREFUL: the slab-emission counter `b` is used directly as the band flat
+   index of the output — here and in the scan walks below.  That identity holds
+   only because the outer iter advances row-major over the complement axes in
+   ascending source order (next_slab_axes in ca_kernel_iterator.c).  If that
+   iteration order changes, every band lands in the wrong output row and
+   nothing raises. */
+
+/* The composite code, data/mask byte offsets and (for min_addr / max_addr) the
+   group-relative source address of a slab element depend only on the slab
+   odometer position, not on the band.  So they are identical for every band.
+   Precompute them once (on the first band, O(slab_elements) = O(group_prod)
+   metadata) and let each band pass just gather the plan + scatter.  This keeps
+   the composite-code bundle walk and the odometer out of the per-(band x
+   element) hot path.  Peak memory stays O(output + group_prod), not O(input). */
+#define GROUP_WALK(T, ACCUM)                                                   \
+  do {                                                                         \
+    ca_iter_state st;                                                          \
+    char       *p;                                                             \
+    boolean8_t *m;                                                             \
+    ca_size_t   b = 0;                                                         \
+    ca_size_t  *gw_ocode = NULL, *gw_doff = NULL, *gw_moff = NULL,             \
+               *gw_gaddr = NULL;                                               \
+    int         gw_ready = 0;                                                  \
+    int         gw_need_addr = ( op == GR_MINADDR || op == GR_MAXADDR );       \
+    CA_FOR_EACH_SLAB(st, ca, axes, (int8_t) ngroup, CA_KERNEL_READ, p, m) {    \
+      int8_t    sndim = st.slab_ndim;                                          \
+      ca_size_t SE    = st.slab_elements;                                      \
+      if ( SE > 0 ) {                                                          \
+        if ( ! gw_ready ) {                                                    \
+          gw_ocode = ALLOC_N(ca_size_t, SE);                                   \
+          gw_doff  = ALLOC_N(ca_size_t, SE);                                   \
+          gw_moff  = ALLOC_N(ca_size_t, SE);                                   \
+          gw_gaddr = ALLOC_N(ca_size_t, SE);                                   \
+          ca_size_t sidx[CA_RANK_MAX];                                         \
+          for ( int8_t k = 0; k < sndim; k++ ) sidx[k] = 0;                    \
+          for ( ca_size_t e = 0; e < SE; e++ ) {                               \
+            /* composite group code from the per-bundle code tables */         \
+            int       skip = 0;                                                \
+            ca_size_t code = 0;                                                \
+            for ( int bi = 0; bi < n_bundles; bi++ ) {                         \
+              ca_size_t sub = 0;                                               \
+              for ( int j = 0; j < bundle_nconsumed[bi]; j++ )                 \
+                sub += sidx[ bundle_slot[bi][j] ] * bundle_cstride[bi][j];     \
+              int32_t c = bundle_codes[bi][sub];                               \
+              if ( c < 0 || (ca_size_t) c >= bundle_k[bi] ) { skip = 1; break; } \
+              code += (ca_size_t) c * bundle_placeval[bi];                     \
+            }                                                                  \
+            ca_size_t doff = 0, moff = 0, gaddr = 0;                           \
+            for ( int8_t k = 0; k < sndim; k++ ) {                             \
+              doff += sidx[k] * st.slab_strides[k];                            \
+              if ( m ) moff += sidx[k] * st.slab_mask_strides[k];             \
+            }                                                                  \
+            if ( gw_need_addr )                                                \
+              for ( long s = 0; s < ngroup; s++ )                             \
+                gaddr += (ca_size_t) sidx[s] * grstride[s];                    \
+            gw_ocode[e] = skip ? GW_SKIP : code * band;                        \
+            gw_doff[e]  = doff;                                                \
+            gw_moff[e]  = moff;                                                \
+            gw_gaddr[e] = gaddr;                                               \
+            /* odometer advance (last slab axis ticks fastest) */             \
+            for ( int8_t k = sndim - 1; k >= 0; k-- ) {                       \
+              if ( ++sidx[k] < st.slab_dims[k] ) break;                       \
+              sidx[k] = 0;                                                     \
+            }                                                                  \
+          }                                                                    \
+          gw_ready = 1;                                                        \
+        }                                                                      \
+        for ( ca_size_t e = 0; e < SE; e++ ) {                                 \
+          ca_size_t oc = gw_ocode[e];                                          \
+          if ( oc == GW_SKIP ) continue;                                       \
+          if ( m && m[ gw_moff[e] ] ) continue;                                \
+          double    v = (double) ( *(T *)(p + gw_doff[e]) );                   \
+          ca_size_t o = oc + b;                                                \
+          ca_size_t gaddr = gw_gaddr[e];                                       \
+          ACCUM;                                                               \
+          (void) v; (void) o; (void) gaddr;                                    \
+        }                                                                      \
+      }                                                                        \
+      b++;                                                                     \
+    }                                                                          \
+    if ( gw_ocode ) xfree(gw_ocode);                                          \
+    if ( gw_doff )  xfree(gw_doff);                                           \
+    if ( gw_moff )  xfree(gw_moff);                                           \
+    if ( gw_gaddr ) xfree(gw_gaddr);                                          \
+  } while (0)
+
+/* Run one walk over every supported native dtype.  Dispatched on the source
+   data_type so the inner loop stays monomorphic (no forced float64 cast). */
+#define GROUP_DISPATCH(ACCUM)                                                  \
+  switch ( ca->data_type ) {                                                   \
+  case CA_BOOLEAN: GROUP_WALK(boolean8_t, ACCUM); break;                       \
+  case CA_INT8:    GROUP_WALK(int8_t,     ACCUM); break;                       \
+  case CA_UINT8:   GROUP_WALK(uint8_t,    ACCUM); break;                       \
+  case CA_INT16:   GROUP_WALK(int16_t,    ACCUM); break;                       \
+  case CA_UINT16:  GROUP_WALK(uint16_t,   ACCUM); break;                       \
+  case CA_INT32:   GROUP_WALK(int32_t,    ACCUM); break;                       \
+  case CA_UINT32:  GROUP_WALK(uint32_t,   ACCUM); break;                       \
+  case CA_INT64:   GROUP_WALK(int64_t,    ACCUM); break;                       \
+  case CA_UINT64:  GROUP_WALK(uint64_t,   ACCUM); break;                       \
+  case CA_FLOAT32: GROUP_WALK(float,      ACCUM); break;                       \
+  case CA_FLOAT64: GROUP_WALK(double,     ACCUM); break;                       \
+  default: break;                                                              \
+  }
+
+/* __axis_group_reduce__(group_axes, bundles, op) — group-reduces self along
+ * the union of `group_axes` (ascending source-axis indices = the slab) into
+ * composite groups described by `bundles`, preserving the band (= non-grouped)
+ * axes.  Internal: the Ruby surface is CAGroupIterator.
+
+   bundles: Array of [codes, k, bundle_axes]
+     - codes       : integer CArray, row-major over `bundle_axes` dims, giving
+                     the group code in [0, k) for each cell of the consumed
+                     axes (rank-1 = per-axis classifier; rank-N = one bundle
+                     consuming >1 source axes = non-rectangular categorical).
+     - k           : Integer, group count for this bundle.
+     - bundle_axes : Array of source-axis indices this bundle consumes
+                     (in the row-major order matching `codes`).
+
+   The bundles' axes together equal `group_axes` (as a set).  Output shape is
+   [K_total, *band_dims] with K_total = Π k over bundles (bundle 0 most
+   significant, row-major) and band_dims = the non-grouped source dims in
+   ascending order.  The Ruby surface reshapes the leading K_total axis into
+   the individual group axes.
+
+   op: :sum :prod :mean :min :max :variance :stddev :variancep :stddevp
+   :count (:count_not_masked is a synonym) :min_addr :max_addr :all :any.
+   :min_addr / :max_addr return the flat raveled source address of the group's
+   extremum, not a group-local index.
+
+   Empty / all-masked groups follow the CArray zero-contribution contract: an
+   identity-bearing op returns its identity (sum 0, prod 1, count 0, all true,
+   any false), a ratio / extremum returns UNDEF.  Sample variance / stddev:
+   n == 0 UNDEF, n == 1 -> 0.0, n >= 2 the formula.
+*/
+static VALUE
+rb_ca_axis_group_reduce (VALUE self, VALUE vgaxes, VALUE vbundles, VALUE vop)
+{
+  CArray *src, *ca, *co;
+  int     op = group_op_code(vop);
+
+  GetCArray(self, src);
+
+  if ( src->ndim <= 0 ) {
+    rb_raise(rb_eRuntimeError, "axis_group_reduce: scalar source");
+  }
+
+  /* --- group (slab) axes --- */
+  Check_Type(vgaxes, T_ARRAY);
+  long ngroup = RARRAY_LEN(vgaxes);
+  if ( ngroup <= 0 || ngroup > src->ndim ) {
+    rb_raise(rb_eArgError, "axis_group_reduce: bad group axis count %ld", ngroup);
+  }
+  int8_t    axes[CA_RANK_MAX];
+  char      is_group[CA_RANK_MAX];
+  for ( int8_t i = 0; i < src->ndim; i++ ) is_group[i] = 0;
+  ca_size_t group_prod = 1;
+  for ( long i = 0; i < ngroup; i++ ) {
+    int a = NUM2INT(RARRAY_AREF(vgaxes, i));
+    if ( a < 0 || a >= src->ndim ) {
+      rb_raise(rb_eArgError, "axis_group_reduce: group axis %d out of range", a);
+    }
+    if ( is_group[a] ) {
+      rb_raise(rb_eArgError, "axis_group_reduce: duplicate group axis %d", a);
+    }
+    if ( i > 0 && a <= NUM2INT(RARRAY_AREF(vgaxes, i - 1)) ) {
+      rb_raise(rb_eArgError, "axis_group_reduce: group axes must be ascending");
+    }
+    is_group[a]  = 1;
+    axes[i]      = (int8_t) a;
+    group_prod  *= src->dim[a];
+  }
+
+  /* --- bundles: small per-group code tables (metadata, kept alive) --- */
+  Check_Type(vbundles, T_ARRAY);
+  int  n_bundles = (int) RARRAY_LEN(vbundles);
+  if ( n_bundles <= 0 || n_bundles > CA_RANK_MAX ) {
+    rb_raise(rb_eArgError, "axis_group_reduce: bad bundle count %d", n_bundles);
+  }
+  int32_t  *bundle_codes[CA_RANK_MAX];
+  ca_size_t bundle_k[CA_RANK_MAX];
+  ca_size_t bundle_placeval[CA_RANK_MAX];
+  int       bundle_nconsumed[CA_RANK_MAX];
+  int       bundle_slot[CA_RANK_MAX][CA_RANK_MAX];
+  ca_size_t bundle_cstride[CA_RANK_MAX][CA_RANK_MAX];
+  CArray   *bundle_ca[CA_RANK_MAX];
+  volatile VALUE keep = rb_ary_new();   /* GC-pin int32 code views */
+
+  ca_size_t K_total = 1;
+  long consumed_total = 0;
+  for ( int bi = 0; bi < n_bundles; bi++ ) {
+    VALUE bundle = RARRAY_AREF(vbundles, bi);
+    Check_Type(bundle, T_ARRAY);
+    if ( RARRAY_LEN(bundle) != 3 ) {
+      rb_raise(rb_eArgError, "axis_group_reduce: bundle must be [codes, k, axes]");
+    }
+    VALUE vcodes = RARRAY_AREF(bundle, 0);
+    ca_size_t k  = (ca_size_t) NUM2LONG(RARRAY_AREF(bundle, 1));
+    VALUE vbaxes = RARRAY_AREF(bundle, 2);
+    Check_Type(vbaxes, T_ARRAY);
+    if ( k <= 0 ) {
+      rb_raise(rb_eArgError, "axis_group_reduce: bundle k must be positive");
+    }
+
+    /* int32 view of the codes (metadata-sized, O(consumed-axis dims)) */
+    VALUE v32 = rb_ca_wrap_readonly(vcodes, INT2NUM(CA_INT32));
+    rb_ary_push((VALUE) keep, v32);
+    GetCArray(v32, bundle_ca[bi]);
+    ca_attach(bundle_ca[bi]);
+    bundle_codes[bi] = (int32_t *) bundle_ca[bi]->ptr;
+
+    int nb = (int) RARRAY_LEN(vbaxes);
+    if ( nb <= 0 || nb > src->ndim ) {
+      rb_raise(rb_eArgError, "axis_group_reduce: bad bundle axis count %d", nb);
+    }
+    bundle_nconsumed[bi] = nb;
+    bundle_k[bi]         = k;
+    K_total             *= k;
+    consumed_total      += nb;
+
+    /* codes must be row-major over the consumed axes' dims */
+    ca_size_t expect = 1;
+    ca_size_t dims[CA_RANK_MAX];
+    for ( int j = 0; j < nb; j++ ) {
+      int a = NUM2INT(RARRAY_AREF(vbaxes, j));
+      if ( a < 0 || a >= src->ndim || ! is_group[a] ) {
+        rb_raise(rb_eArgError,
+                 "axis_group_reduce: bundle axis %d not a group axis", a);
+      }
+      dims[j]  = src->dim[a];
+      expect  *= src->dim[a];
+      /* slot = position of source axis `a` within the ascending slab axes */
+      int slot = -1;
+      for ( long s = 0; s < ngroup; s++ ) {
+        if ( axes[s] == a ) { slot = (int) s; break; }
+      }
+      bundle_slot[bi][j] = slot;   /* always found: a is a group axis */
+    }
+    if ( bundle_ca[bi]->elements != expect ) {
+      rb_raise(rb_eArgError,
+               "axis_group_reduce: codes length %lld != Π consumed dims %lld",
+               (long long) bundle_ca[bi]->elements, (long long) expect);
+    }
+    /* row-major code strides over the consumed axes (in given order) */
+    ca_size_t st = 1;
+    for ( int j = nb - 1; j >= 0; j-- ) {
+      bundle_cstride[bi][j] = st;
+      st *= dims[j];
+    }
+  }
+  if ( consumed_total != ngroup ) {
+    rb_raise(rb_eArgError,
+             "axis_group_reduce: Σ bundle ranks %ld != group axis count %ld",
+             consumed_total, ngroup);
+  }
+  /* place values: bundle 0 most significant (row-major over bundles) */
+  {
+    ca_size_t pv = 1;
+    for ( int bi = n_bundles - 1; bi >= 0; bi-- ) {
+      bundle_placeval[bi] = pv;
+      pv *= bundle_k[bi];
+    }
+  }
+
+  /* --- band layout + output shape [K_total, *band_dims] --- */
+  ca_size_t band = (group_prod > 0) ? (src->elements / group_prod) : 0;
+  ca_size_t odim[CA_RANK_MAX];
+  int8_t    ondim = 1;
+  odim[0] = K_total;
+  for ( int8_t i = 0; i < src->ndim; i++ ) {
+    if ( ! is_group[i] ) odim[ondim++] = src->dim[i];
+  }
+  ca_size_t nout = K_total * band;
+
+  /* --- supported dtype gate (before any allocation) --- */
+  ca = src;
+  switch ( src->data_type ) {
+  case CA_BOOLEAN: case CA_INT8:  case CA_UINT8:  case CA_INT16: case CA_UINT16:
+  case CA_INT32:   case CA_UINT32: case CA_INT64: case CA_UINT64:
+  case CA_FLOAT32: case CA_FLOAT64: break;
+  default:
+    for ( int bi = 0; bi < n_bundles; bi++ ) ca_detach(bundle_ca[bi]);
+    rb_raise(rb_eRuntimeError,
+             "axis_group_reduce: unsupported source data_type %d",
+             src->data_type);
+  }
+
+  /* output dtype per op */
+  int8_t out_dt = CA_FLOAT64;
+  if      ( op == GR_COUNT )              out_dt = CA_INT64;
+  else if ( op == GR_MINADDR || op == GR_MAXADDR ) out_dt = CA_INT64;
+  else if ( op == GR_ALL || op == GR_ANY ) out_dt = CA_BOOLEAN;
+  VALUE vout = rb_carray_new(out_dt, ondim, odim, 0, NULL);
+  GetCArray(vout, co);
+
+  /* --- accumulator buffers (only those the op needs; all O(nout)) --- */
+  ca_size_t *cnt   = ALLOC_N(ca_size_t, nout);  MEMZERO(cnt, ca_size_t, nout);
+  double    *sum   = NULL, *sumsq = NULL, *prod = NULL, *mn = NULL, *mx = NULL;
+  ca_size_t *nz    = NULL;
+  int64_t   *mnaddr = NULL, *mxaddr = NULL;   /* flat source addr of min / max */
+  ca_size_t *band_addr = NULL;                /* raveled addr of each band cell */
+  if ( op == GR_SUM || op == GR_MEAN || op == GR_VARIANCE || op == GR_STDDEV ||
+       op == GR_VARIANCEP || op == GR_STDDEVP ) {
+    sum = ALLOC_N(double, nout);  MEMZERO(sum, double, nout);
+  }
+  if ( op == GR_PROD ) {
+    prod = ALLOC_N(double, nout);
+    for ( ca_size_t o = 0; o < nout; o++ ) prod[o] = 1.0;
+  }
+  if ( op == GR_MIN || op == GR_MINADDR ) {
+    mn = ALLOC_N(double, nout);
+    for ( ca_size_t o = 0; o < nout; o++ ) mn[o] = HUGE_VAL;
+  }
+  if ( op == GR_MAX || op == GR_MAXADDR ) {
+    mx = ALLOC_N(double, nout);
+    for ( ca_size_t o = 0; o < nout; o++ ) mx[o] = -HUGE_VAL;
+  }
+  if ( op == GR_MINADDR ) { mnaddr = ALLOC_N(int64_t, nout); MEMZERO(mnaddr, int64_t, nout); }
+  if ( op == GR_MAXADDR ) { mxaddr = ALLOC_N(int64_t, nout); MEMZERO(mxaddr, int64_t, nout); }
+  if ( op == GR_ALL || op == GR_ANY ) {
+    nz = ALLOC_N(ca_size_t, nout);  MEMZERO(nz, ca_size_t, nout);
+  }
+
+  /* min_addr / max_addr need the flat raveled source address of each cell.
+     Precompute the raveled stride per axis, the group-axis strides (paired with
+     the slab odometer sidx), and the band cell's base address per band flat
+     index b. Then addr(cell) = band_addr[b] + Σ sidx[s]*grstride[s]. */
+  ca_size_t rstride[CA_RANK_MAX], grstride[CA_RANK_MAX];
+  if ( op == GR_MINADDR || op == GR_MAXADDR ) {
+    rstride[src->ndim - 1] = 1;
+    for ( int8_t k = (int8_t)(src->ndim - 2); k >= 0; k-- )
+      rstride[k] = rstride[k+1] * src->dim[k+1];
+    for ( long s = 0; s < ngroup; s++ ) grstride[s] = rstride[axes[s]];
+    int band_axis[CA_RANK_MAX]; int nband_axes = 0;
+    for ( int8_t k = 0; k < src->ndim; k++ )
+      if ( ! is_group[k] ) band_axis[nband_axes++] = k;
+    band_addr = ALLOC_N(ca_size_t, band > 0 ? band : 1);
+    for ( ca_size_t bb = 0; bb < band; bb++ ) {
+      ca_size_t rem = bb, addr = 0;
+      for ( int j = nband_axes - 1; j >= 0; j-- ) {     /* last band axis fastest */
+        ca_size_t d = src->dim[band_axis[j]];
+        addr += (rem % d) * rstride[band_axis[j]];
+        rem  /= d;
+      }
+      band_addr[bb] = addr;
+    }
+  }
+
+  /* --- compute pass(es), native dtype dispatch, no forced float64 cast ---
+     variance / stddev use a centred two-pass (= matches CArray's own
+     variance, avoids the one-pass sumsq cancellation that breaks ε-close
+     for small near-constant groups).  Pass 1 fills sum + cnt; sum is then
+     overwritten in place with the per-group mean; pass 2 accumulates the
+     centred sum of squares.  Every other op is a single pass. */
+  if ( op == GR_VARIANCE || op == GR_STDDEV ||
+       op == GR_VARIANCEP || op == GR_STDDEVP ) {
+    GROUP_DISPATCH( cnt[o] += 1; sum[o] += v; );
+    for ( ca_size_t o = 0; o < nout; o++ )
+      if ( cnt[o] > 0 ) sum[o] /= (double) cnt[o];   /* sum -> mean */
+    sumsq = ALLOC_N(double, nout);  MEMZERO(sumsq, double, nout);
+    GROUP_DISPATCH( { double _d = v - sum[o]; sumsq[o] += _d * _d; } );
+  }
+  else if ( op == GR_MINADDR ) {
+    GROUP_DISPATCH(
+      cnt[o] += 1;
+      if ( v < mn[o] ) {
+        ca_size_t faddr = band_addr[b] + gaddr;
+        mn[o] = v; mnaddr[o] = (int64_t) faddr;
+      }
+    );
+  }
+  else if ( op == GR_MAXADDR ) {
+    GROUP_DISPATCH(
+      cnt[o] += 1;
+      if ( v > mx[o] ) {
+        ca_size_t faddr = band_addr[b] + gaddr;
+        mx[o] = v; mxaddr[o] = (int64_t) faddr;
+      }
+    );
+  }
+  else {
+    GROUP_DISPATCH(
+      cnt[o] += 1;
+      if ( sum )  sum[o]  += v;
+      if ( prod ) prod[o] *= v;
+      if ( mn )   { if ( v < mn[o] ) mn[o] = v; }
+      if ( mx )   { if ( v > mx[o] ) mx[o] = v; }
+      if ( nz )   { if ( v != 0.0 ) nz[o] += 1; }
+    );
+  }
+
+  /* --- finalise into the output, UNDEF for empty groups --- */
+  boolean8_t *omask = NULL;
+  #define MARK_UNDEF(o) do {                       \
+      if ( ! omask ) {                             \
+        ca_create_mask(co);                        \
+        omask = (boolean8_t *) co->mask->ptr;      \
+      }                                            \
+      omask[o] = 1;                                \
+    } while (0)
+
+  /* Empty / all-masked groups follow the same zero-contribution contract as
+     CArray reductions (ERI, matching the categorical sibling): an identity-
+     bearing op returns its identity (sum 0, prod 1, count 0, all true, any
+     false), a ratio / extremum returns UNDEF. Sample variance/stddev: n==0
+     UNDEF, n==1 -> 0.0 (the n=1 contract), n>=2 the formula. */
+  if ( op == GR_COUNT ) {
+    int64_t *out = (int64_t *) co->ptr;
+    for ( ca_size_t o = 0; o < nout; o++ ) out[o] = (int64_t) cnt[o];
+  }
+  else if ( op == GR_MINADDR || op == GR_MAXADDR ) {
+    int64_t *out  = (int64_t *) co->ptr;
+    int64_t *addr = ( op == GR_MINADDR ) ? mnaddr : mxaddr;
+    for ( ca_size_t o = 0; o < nout; o++ ) {
+      if ( cnt[o] == 0 ) { out[o] = 0; MARK_UNDEF(o); }   /* empty -> UNDEF */
+      else out[o] = addr[o];
+    }
+  }
+  else if ( op == GR_ALL ) {
+    boolean8_t *out = (boolean8_t *) co->ptr;    /* empty -> true (vacuous) */
+    for ( ca_size_t o = 0; o < nout; o++ )
+      out[o] = ( cnt[o] == 0 ) ? 1 : (nz[o] == cnt[o]);
+  }
+  else if ( op == GR_ANY ) {
+    boolean8_t *out = (boolean8_t *) co->ptr;    /* empty -> false */
+    for ( ca_size_t o = 0; o < nout; o++ )
+      out[o] = ( cnt[o] == 0 ) ? 0 : (nz[o] > 0);
+  }
+  else {
+    double *out = (double *) co->ptr;
+    for ( ca_size_t o = 0; o < nout; o++ ) {
+      switch ( op ) {
+      case GR_SUM:  out[o] = sum[o];  break;   /* empty -> 0.0 (identity) */
+      case GR_PROD: out[o] = prod[o]; break;   /* empty -> 1.0 (identity) */
+      case GR_MEAN:
+        if ( cnt[o] == 0 ) { out[o] = 0.0; MARK_UNDEF(o); }
+        else out[o] = sum[o] / (double) cnt[o];
+        break;
+      case GR_MIN:
+        if ( cnt[o] == 0 ) { out[o] = 0.0; MARK_UNDEF(o); } else out[o] = mn[o];
+        break;
+      case GR_MAX:
+        if ( cnt[o] == 0 ) { out[o] = 0.0; MARK_UNDEF(o); } else out[o] = mx[o];
+        break;
+      case GR_VARIANCE:
+      case GR_STDDEV:
+        if ( cnt[o] == 0 ) { out[o] = 0.0; MARK_UNDEF(o); }
+        else if ( cnt[o] == 1 ) { out[o] = 0.0; }         /* n=1 contract */
+        else { double var = sumsq[o] / ( (double) cnt[o] - 1.0 );
+               out[o] = ( op == GR_STDDEV ) ? sqrt(var) : var; }
+        break;
+      case GR_VARIANCEP:
+      case GR_STDDEVP:
+        if ( cnt[o] == 0 ) { out[o] = 0.0; MARK_UNDEF(o); }
+        else { double varp = sumsq[o] / (double) cnt[o];   /* n>=1, n=1 -> 0.0 */
+               out[o] = ( op == GR_STDDEVP ) ? sqrt(varp) : varp; }
+        break;
+      }
+    }
+  }
+  #undef MARK_UNDEF
+
+  for ( int bi = 0; bi < n_bundles; bi++ ) ca_detach(bundle_ca[bi]);
+  xfree(cnt);
+  if ( sum ) xfree(sum);
+  if ( sumsq ) xfree(sumsq);
+  if ( prod ) xfree(prod);
+  if ( mn ) xfree(mn);
+  if ( mx ) xfree(mx);
+  if ( nz ) xfree(nz);
+  if ( mnaddr ) xfree(mnaddr);
+  if ( mxaddr ) xfree(mxaddr);
+  if ( band_addr ) xfree(band_addr);
+
+  RB_GC_GUARD(keep);
+  return vout;
+}
+
+/* =========================================================================
+   Grouped segment scan — the per-element-emit sibling of the reduce above.
+
+   Same fused single pass (CA_FOR_EACH_SLAB pins the grouped-axis union as the
+   slab, the band axes as the outer iter, composite group code per slab element
+   from the bundles), but instead of reading a dense accumulator at the end it
+   EMITS the running value per element into a source-shaped output:
+
+     reduce:  acc[code] op= v            then out[code, band] = acc[code]
+     scan:    acc[code] op= v ; out[cell] = acc[code]   (shape = source)
+
+   The group axis is NOT collapsed — output shape == source shape. Within each
+   band the slab elements are walked in odometer (row-major over the grouped
+   axes) order, so per group the accumulation runs in position order along the
+   grouped axes. The flat case (all axes grouped, one band) is the same walk
+   with a single slab = row-major appearance order, so flat and band-preserving
+   agree by construction.
+
+   The accumulator is dense [0, K_total) and is RESET at each band (each slab
+   is one band position, fully processed before the next), so peak extra memory
+   is O(K_total) — never O(input). Peak = output (= input size) + O(K_total).
+
+   Masked / excluded cells follow the core CArray scan: a slab element masked in
+   the source but with a valid group code does not update its group's total; it
+   HOLDS the current running value and its output is NOT masked (identity ops
+   always; extrema only once a member has been seen, else UNDEF — the empty
+   max/min reduction contract).  A slab element excluded by the classifier (code
+   out of [0, k)) belongs to no group and stays UNDEF.
+   ------------------------------------------------------------------------- */
+
+enum { GS_CUMSUM = 0, GS_CUMCOUNT, GS_CUMMAX, GS_CUMMIN, GS_CUMPROD };
+
+/* Build the band-independent scatter plan for one slab.  For each slab element
+   (in odometer order over the grouped axes) it records: the composite group
+   code (GW_SKIP if excluded by some bundle), the source data / mask byte
+   offsets from the slab pointers, and the group-relative raveled output address
+   (Σ sidx[s]*grstride[s]).  These depend only on the slab odometer position,
+   not on the band, so every band reuses this one plan (built on the first
+   band).  Keeps the bundle walk + odometer out of the per-(band × element) hot
+   path; peak metadata is O(slab_elements) = O(group_prod), never O(input). */
+static void
+group_scan_build_plan (ca_iter_state *st, boolean8_t *m,
+                       int n_bundles, int32_t **bundle_codes,
+                       ca_size_t *bundle_k, ca_size_t *bundle_placeval,
+                       int *bundle_nconsumed, int (*bundle_slot)[CA_RANK_MAX],
+                       ca_size_t (*bundle_cstride)[CA_RANK_MAX],
+                       long ngroup, ca_size_t *grstride,
+                       ca_size_t *sw_code, ca_size_t *sw_doff,
+                       ca_size_t *sw_moff, ca_size_t *sw_addr)
+{
+  int8_t    sndim = st->slab_ndim;
+  ca_size_t SE    = st->slab_elements;
+  ca_size_t sidx[CA_RANK_MAX];
+  for ( int8_t k = 0; k < sndim; k++ ) {
+    sidx[k] = 0;
+  }
+  for ( ca_size_t e = 0; e < SE; e++ ) {
+    int       skip = 0;
+    ca_size_t code = 0;
+    for ( int bi = 0; bi < n_bundles; bi++ ) {
+      ca_size_t sub = 0;
+      for ( int j = 0; j < bundle_nconsumed[bi]; j++ ) {
+        sub += sidx[ bundle_slot[bi][j] ] * bundle_cstride[bi][j];
+      }
+      int32_t c = bundle_codes[bi][sub];
+      if ( c < 0 || (ca_size_t) c >= bundle_k[bi] ) {
+        skip = 1;
+        break;
+      }
+      code += (ca_size_t) c * bundle_placeval[bi];
+    }
+    ca_size_t doff = 0, moff = 0, gaddr = 0;
+    for ( int8_t k = 0; k < sndim; k++ ) {
+      doff += sidx[k] * st->slab_strides[k];
+      if ( m ) {
+        moff += sidx[k] * st->slab_mask_strides[k];
+      }
+    }
+    for ( long s = 0; s < ngroup; s++ ) {
+      gaddr += (ca_size_t) sidx[s] * grstride[s];
+    }
+    sw_code[e] = skip ? GW_SKIP : code;
+    sw_doff[e] = doff;
+    sw_moff[e] = moff;
+    sw_addr[e] = gaddr;
+    for ( int8_t k = sndim - 1; k >= 0; k-- ) {   /* last slab axis ticks fastest */
+      if ( ++sidx[k] < st->slab_dims[k] ) {
+        break;
+      }
+      sidx[k] = 0;
+    }
+  }
+}
+
+/* GROUP_SCAN_WALK(T, EMIT, HOLD): the accumulate-in-double walk (cumsum /
+   cumprod / cumcount).  One CA_FOR_EACH_SLAB pass monomorphised on the load
+   type T; the band-independent plan is built once (group_scan_build_plan) on
+   the first band.  EMIT (a present cell) writes the running value at the cell's
+   raveled output address `addr`, with `v` (the element widened to double),
+   `code` (composite group code) and the accumulators accd / accn in scope.
+   These ops have an identity (0.0 sum / 1.0 prod / 0 count), so the accumulator
+   is valid before any member: a cell masked WITHIN its group holds the current
+   running value (HOLD) and its output is NOT masked — it just does not update
+   the accumulator, matching the core scan where a masked cell holds the running
+   acc.  Only a cell excluded from every group (GW_SKIP) stays UNDEF.  At each
+   band accd is reset to acc_init and accn to 0, so groups do not run across
+   bands. */
+#define GROUP_SCAN_WALK(T, EMIT, HOLD)                                        \
+  do {                                                                        \
+    ca_iter_state st;                                                         \
+    char       *p;                                                            \
+    boolean8_t *m;                                                            \
+    ca_size_t   b = 0;                                                        \
+    int         ready = 0;                                                    \
+    CA_FOR_EACH_SLAB(st, ca, axes, (int8_t) ngroup, CA_KERNEL_READ, p, m) {   \
+      ca_size_t SE = st.slab_elements;                                        \
+      if ( SE > 0 ) {                                                         \
+        if ( ! ready ) {                                                      \
+          group_scan_build_plan(&st, m, n_bundles, bundle_codes, bundle_k,   \
+                                bundle_placeval, bundle_nconsumed,            \
+                                bundle_slot, bundle_cstride, ngroup,          \
+                                grstride, sw_code, sw_doff, sw_moff, sw_addr);\
+          ready = 1;                                                          \
+        }                                                                     \
+        for ( ca_size_t o = 0; o < K_total; o++ ) {                          \
+          accd[o] = acc_init;                                                 \
+          accn[o] = 0;                                                        \
+        }                                                                     \
+        ca_size_t base = band_addr[b];                                       \
+        for ( ca_size_t e = 0; e < SE; e++ ) {                               \
+          ca_size_t code = sw_code[e];                                        \
+          ca_size_t addr = base + sw_addr[e];                                 \
+          if ( code == GW_SKIP ) { MARK_OUT_UNDEF(addr); continue; }          \
+          if ( m && m[ sw_moff[e] ] ) { HOLD; continue; }                     \
+          double v = (double) ( *(T *)(p + sw_doff[e]) );                     \
+          (void) v;                                                           \
+          EMIT;                                                               \
+        }                                                                     \
+      }                                                                       \
+      b++;                                                                    \
+    }                                                                         \
+  } while (0)
+
+#define GROUP_SCAN_DISPATCH(EMIT, HOLD)                                       \
+  switch ( ca->data_type ) {                                                  \
+  case CA_BOOLEAN: GROUP_SCAN_WALK(boolean8_t, EMIT, HOLD); break;            \
+  case CA_INT8:    GROUP_SCAN_WALK(int8_t,     EMIT, HOLD); break;            \
+  case CA_UINT8:   GROUP_SCAN_WALK(uint8_t,    EMIT, HOLD); break;            \
+  case CA_INT16:   GROUP_SCAN_WALK(int16_t,    EMIT, HOLD); break;            \
+  case CA_UINT16:  GROUP_SCAN_WALK(uint16_t,   EMIT, HOLD); break;            \
+  case CA_INT32:   GROUP_SCAN_WALK(int32_t,    EMIT, HOLD); break;            \
+  case CA_UINT32:  GROUP_SCAN_WALK(uint32_t,   EMIT, HOLD); break;            \
+  case CA_INT64:   GROUP_SCAN_WALK(int64_t,    EMIT, HOLD); break;            \
+  case CA_UINT64:  GROUP_SCAN_WALK(uint64_t,   EMIT, HOLD); break;            \
+  case CA_FLOAT32: GROUP_SCAN_WALK(float,      EMIT, HOLD); break;            \
+  case CA_FLOAT64: GROUP_SCAN_WALK(double,     EMIT, HOLD); break;            \
+  default: break;                                                             \
+  }
+
+/* GROUP_SCAN_EXTREMUM_WALK(T, CMP): running extremum (cummax / cummin).  The
+   extremum keeps the source dtype (its magnitude never grows), so it holds a
+   native T accumulator, not a widened double.  The first member of a group
+   emits its own value: a per-group `seen` byte initialises the accumulator
+   lazily on first hit — no sentinel like HUGE_VAL, which an integer dtype could
+   not represent.  CMP is > for max, < for min: a later member replaces the
+   running extremum when `rv CMP acc`.  A cell masked within its group holds the
+   current extremum once a member has been seen (output NOT masked, like sum);
+   before any member the extremum is undefined so the cell stays UNDEF (the
+   CArray reduction contract: empty max/min has no value — deliberately NOT the
+   type-min/-Inf init the core cumulative uses).  GW_SKIP (excluded) stays
+   UNDEF. */
+#define GROUP_SCAN_EXTREMUM_WALK(T, CMP)                                      \
+  do {                                                                        \
+    ca_iter_state st;                                                         \
+    char       *p;                                                            \
+    boolean8_t *m;                                                            \
+    ca_size_t   b = 0;                                                        \
+    int         ready = 0;                                                    \
+    T          *acce = ALLOC_N(T, K_total);                                   \
+    T          *outp = (T *) co->ptr;                                         \
+    CA_FOR_EACH_SLAB(st, ca, axes, (int8_t) ngroup, CA_KERNEL_READ, p, m) {   \
+      ca_size_t SE = st.slab_elements;                                        \
+      if ( SE > 0 ) {                                                         \
+        if ( ! ready ) {                                                      \
+          group_scan_build_plan(&st, m, n_bundles, bundle_codes, bundle_k,   \
+                                bundle_placeval, bundle_nconsumed,            \
+                                bundle_slot, bundle_cstride, ngroup,          \
+                                grstride, sw_code, sw_doff, sw_moff, sw_addr);\
+          ready = 1;                                                          \
+        }                                                                     \
+        for ( ca_size_t o = 0; o < K_total; o++ ) {                          \
+          seen[o] = 0;                                                        \
+        }                                                                     \
+        ca_size_t base = band_addr[b];                                       \
+        for ( ca_size_t e = 0; e < SE; e++ ) {                               \
+          ca_size_t code = sw_code[e];                                        \
+          ca_size_t addr = base + sw_addr[e];                                 \
+          if ( code == GW_SKIP ) { MARK_OUT_UNDEF(addr); continue; }          \
+          if ( m && m[ sw_moff[e] ] ) {                                       \
+            if ( seen[code] ) { outp[addr] = acce[code]; }                    \
+            else { MARK_OUT_UNDEF(addr); }                                    \
+            continue;                                                         \
+          }                                                                   \
+          T rv = *(T *)(p + sw_doff[e]);                                      \
+          if ( ! seen[code] ) { acce[code] = rv; seen[code] = 1; }            \
+          else if ( rv CMP acce[code] ) { acce[code] = rv; }                  \
+          outp[addr] = acce[code];                                            \
+        }                                                                     \
+      }                                                                       \
+      b++;                                                                    \
+    }                                                                         \
+    xfree(acce);                                                              \
+  } while (0)
+
+#define GROUP_SCAN_EXTREMUM_DISPATCH(CMP)                                     \
+  switch ( ca->data_type ) {                                                  \
+  case CA_BOOLEAN: GROUP_SCAN_EXTREMUM_WALK(boolean8_t, CMP); break;          \
+  case CA_INT8:    GROUP_SCAN_EXTREMUM_WALK(int8_t,     CMP); break;          \
+  case CA_UINT8:   GROUP_SCAN_EXTREMUM_WALK(uint8_t,    CMP); break;          \
+  case CA_INT16:   GROUP_SCAN_EXTREMUM_WALK(int16_t,    CMP); break;          \
+  case CA_UINT16:  GROUP_SCAN_EXTREMUM_WALK(uint16_t,   CMP); break;          \
+  case CA_INT32:   GROUP_SCAN_EXTREMUM_WALK(int32_t,    CMP); break;          \
+  case CA_UINT32:  GROUP_SCAN_EXTREMUM_WALK(uint32_t,   CMP); break;          \
+  case CA_INT64:   GROUP_SCAN_EXTREMUM_WALK(int64_t,    CMP); break;          \
+  case CA_UINT64:  GROUP_SCAN_EXTREMUM_WALK(uint64_t,   CMP); break;          \
+  case CA_FLOAT32: GROUP_SCAN_EXTREMUM_WALK(float,      CMP); break;          \
+  case CA_FLOAT64: GROUP_SCAN_EXTREMUM_WALK(double,     CMP); break;          \
+  default: break;                                                             \
+  }
+
+/* GROUP_SCAN_OBJECT_WALK(EMIT, HOLD): the CA_OBJECT lane (source holds VALUEs).
+   A per-cell rb_funcall walk: cumcount is an int64 running count; the arithmetic
+   ops (cumsum / cumprod) accumulate a per-group VALUE from an identity seed (0 /
+   1) and the extremum ops (cummax / cummin) from the first member itself, all
+   emitting into a CA_OBJECT output.  EMIT sees `ev` (the source element VALUE)
+   and writes each running acc to the output cell immediately after computing it,
+   so between rb_funcall and the store no allocation runs and every live
+   accumulator is reachable from co (GC-safe).
+   A cell masked within its group runs HOLD, not the emit: cumcount holds the
+   running count and cumsum / cumprod hold their identity-seeded acc (0 / 1),
+   both unmasked (matching core object cumsum / cumprod); cummax / cummin hold
+   the running VALUE only once a member has been seen and stay UNDEF before that
+   (no identity for arbitrary objects).  GW_SKIP (excluded) stays UNDEF.  accn /
+   seen are reset at each band; acco carries the running VALUE. */
+#define GROUP_SCAN_OBJECT_WALK(EMIT, HOLD)                                    \
+  do {                                                                        \
+    ca_iter_state st;                                                         \
+    char       *p;                                                            \
+    boolean8_t *m;                                                            \
+    ca_size_t   b = 0;                                                        \
+    int         ready = 0;                                                    \
+    CA_FOR_EACH_SLAB(st, ca, axes, (int8_t) ngroup, CA_KERNEL_READ, p, m) {   \
+      ca_size_t SE = st.slab_elements;                                        \
+      if ( SE > 0 ) {                                                         \
+        if ( ! ready ) {                                                      \
+          group_scan_build_plan(&st, m, n_bundles, bundle_codes, bundle_k,   \
+                                bundle_placeval, bundle_nconsumed,            \
+                                bundle_slot, bundle_cstride, ngroup,          \
+                                grstride, sw_code, sw_doff, sw_moff, sw_addr);\
+          ready = 1;                                                          \
+        }                                                                     \
+        for ( ca_size_t o = 0; o < K_total; o++ ) {                          \
+          seen[o] = 0;                                                        \
+          accn[o] = 0;                                                        \
+        }                                                                     \
+        ca_size_t base = band_addr[b];                                       \
+        for ( ca_size_t e = 0; e < SE; e++ ) {                               \
+          ca_size_t code = sw_code[e];                                        \
+          ca_size_t addr = base + sw_addr[e];                                 \
+          if ( code == GW_SKIP ) { MARK_OUT_UNDEF(addr); continue; }          \
+          if ( m && m[ sw_moff[e] ] ) { HOLD; continue; }                     \
+          VALUE ev = *(VALUE *)(p + sw_doff[e]);                              \
+          (void) ev;                                                          \
+          EMIT;                                                               \
+        }                                                                     \
+      }                                                                       \
+      b++;                                                                    \
+    }                                                                         \
+  } while (0)
+
+static int
+group_scan_op_code (VALUE vop)
+{
+  ID id = SYM2ID(vop);
+  if      ( id == rb_intern("cumsum") )   return GS_CUMSUM;
+  else if ( id == rb_intern("cumcount") ) return GS_CUMCOUNT;
+  else if ( id == rb_intern("cummax") )   return GS_CUMMAX;
+  else if ( id == rb_intern("cummin") )   return GS_CUMMIN;
+  else if ( id == rb_intern("cumprod") )  return GS_CUMPROD;
+  rb_raise(rb_eArgError, "axis_group_scan: unsupported op :%s", rb_id2name(id));
+}
+
+/* __axis_group_scan__(group_axes, bundles, op) — group-keyed segment scan of
+ * self over the union of `group_axes` into composite groups described by
+ * `bundles` (same argument shape as __axis_group_reduce__).  Internal.
+ *
+ * Unlike the reduce the grouped axes are not collapsed: the result is a new
+ * CArray of the same shape as self, each cell holding the running statistic of
+ * its group up to and including that cell, in row-major position order along
+ * the grouped axes (per band).
+
+   op / output dtype:
+     :cumsum   -> float64, inclusive within-group running sum.
+     :cumprod  -> float64, inclusive within-group running product (init 1.0;
+                  float64 like cumsum since the product grows).
+     :cummax   -> source dtype, running within-group maximum (extrema do not
+                  grow magnitude, so the dtype is preserved; int stays int).
+     :cummin   -> source dtype, running within-group minimum.
+     :cumcount -> int64, 1-based within-group running count of present cells
+                  (matching the core CArray#cumcount): the first present member
+                  of a group emits 1, the next 2, ...
+   cumsum / cumprod keep float64 (matching the reduce siblings sum / prod);
+   integer-preserving sum / prod is a deliberate non-goal (overflow / dtype
+   consistency), as on the reduce side.  A CA_OBJECT source emits a CA_OBJECT
+   result for cumsum / cumprod / cummax / cummin (cumcount stays int64).
+
+   Masked-cell / excluded-cell policy (matches the core CArray scan): a cell
+   masked WITHIN its group holds its group's current running value and its
+   output is NOT masked — for cumsum / cumprod / cumcount (which have an
+   identity) always, for cummax / cummin (no identity) only once a member has
+   been seen; before any member the extremum is undefined so the cell stays
+   UNDEF.  A cell excluded from every group (composite code out of [0,k)) belongs
+   to no group and stays UNDEF. */
+static VALUE
+rb_ca_axis_group_scan (VALUE self, VALUE vgaxes, VALUE vbundles, VALUE vop)
+{
+  CArray *src, *ca, *co;
+  int     op = group_scan_op_code(vop);
+
+  GetCArray(self, src);
+
+  if ( src->ndim <= 0 ) {
+    rb_raise(rb_eRuntimeError, "axis_group_scan: scalar source");
+  }
+
+  /* --- group (slab) axes (same validation as reduce) --- */
+  Check_Type(vgaxes, T_ARRAY);
+  long ngroup = RARRAY_LEN(vgaxes);
+  if ( ngroup <= 0 || ngroup > src->ndim ) {
+    rb_raise(rb_eArgError, "axis_group_scan: bad group axis count %ld", ngroup);
+  }
+  int8_t    axes[CA_RANK_MAX];
+  char      is_group[CA_RANK_MAX];
+  for ( int8_t i = 0; i < src->ndim; i++ ) is_group[i] = 0;
+  ca_size_t group_prod = 1;
+  for ( long i = 0; i < ngroup; i++ ) {
+    int a = NUM2INT(RARRAY_AREF(vgaxes, i));
+    if ( a < 0 || a >= src->ndim ) {
+      rb_raise(rb_eArgError, "axis_group_scan: group axis %d out of range", a);
+    }
+    if ( is_group[a] ) {
+      rb_raise(rb_eArgError, "axis_group_scan: duplicate group axis %d", a);
+    }
+    if ( i > 0 && a <= NUM2INT(RARRAY_AREF(vgaxes, i - 1)) ) {
+      rb_raise(rb_eArgError, "axis_group_scan: group axes must be ascending");
+    }
+    is_group[a]  = 1;
+    axes[i]      = (int8_t) a;
+    group_prod  *= src->dim[a];
+  }
+
+  /* --- bundles: small per-group code tables (same as reduce) --- */
+  Check_Type(vbundles, T_ARRAY);
+  int  n_bundles = (int) RARRAY_LEN(vbundles);
+  if ( n_bundles <= 0 || n_bundles > CA_RANK_MAX ) {
+    rb_raise(rb_eArgError, "axis_group_scan: bad bundle count %d", n_bundles);
+  }
+  int32_t  *bundle_codes[CA_RANK_MAX];
+  ca_size_t bundle_k[CA_RANK_MAX];
+  ca_size_t bundle_placeval[CA_RANK_MAX];
+  int       bundle_nconsumed[CA_RANK_MAX];
+  int       bundle_slot[CA_RANK_MAX][CA_RANK_MAX];
+  ca_size_t bundle_cstride[CA_RANK_MAX][CA_RANK_MAX];
+  CArray   *bundle_ca[CA_RANK_MAX];
+  volatile VALUE keep = rb_ary_new();
+
+  ca_size_t K_total = 1;
+  long consumed_total = 0;
+  for ( int bi = 0; bi < n_bundles; bi++ ) {
+    VALUE bundle = RARRAY_AREF(vbundles, bi);
+    Check_Type(bundle, T_ARRAY);
+    if ( RARRAY_LEN(bundle) != 3 ) {
+      rb_raise(rb_eArgError, "axis_group_scan: bundle must be [codes, k, axes]");
+    }
+    VALUE vcodes = RARRAY_AREF(bundle, 0);
+    ca_size_t k  = (ca_size_t) NUM2LONG(RARRAY_AREF(bundle, 1));
+    VALUE vbaxes = RARRAY_AREF(bundle, 2);
+    Check_Type(vbaxes, T_ARRAY);
+    if ( k <= 0 ) {
+      rb_raise(rb_eArgError, "axis_group_scan: bundle k must be positive");
+    }
+
+    VALUE v32 = rb_ca_wrap_readonly(vcodes, INT2NUM(CA_INT32));
+    rb_ary_push((VALUE) keep, v32);
+    GetCArray(v32, bundle_ca[bi]);
+    ca_attach(bundle_ca[bi]);
+    bundle_codes[bi] = (int32_t *) bundle_ca[bi]->ptr;
+
+    int nb = (int) RARRAY_LEN(vbaxes);
+    if ( nb <= 0 || nb > src->ndim ) {
+      rb_raise(rb_eArgError, "axis_group_scan: bad bundle axis count %d", nb);
+    }
+    bundle_nconsumed[bi] = nb;
+    bundle_k[bi]         = k;
+    K_total             *= k;
+    consumed_total      += nb;
+
+    ca_size_t expect = 1;
+    ca_size_t dims[CA_RANK_MAX];
+    for ( int j = 0; j < nb; j++ ) {
+      int a = NUM2INT(RARRAY_AREF(vbaxes, j));
+      if ( a < 0 || a >= src->ndim || ! is_group[a] ) {
+        rb_raise(rb_eArgError,
+                 "axis_group_scan: bundle axis %d not a group axis", a);
+      }
+      dims[j]  = src->dim[a];
+      expect  *= src->dim[a];
+      int slot = -1;
+      for ( long s = 0; s < ngroup; s++ ) {
+        if ( axes[s] == a ) { slot = (int) s; break; }
+      }
+      bundle_slot[bi][j] = slot;
+    }
+    if ( bundle_ca[bi]->elements != expect ) {
+      rb_raise(rb_eArgError,
+               "axis_group_scan: codes length %lld != Π consumed dims %lld",
+               (long long) bundle_ca[bi]->elements, (long long) expect);
+    }
+    ca_size_t cs = 1;
+    for ( int j = nb - 1; j >= 0; j-- ) {
+      bundle_cstride[bi][j] = cs;
+      cs *= dims[j];
+    }
+  }
+  if ( consumed_total != ngroup ) {
+    rb_raise(rb_eArgError,
+             "axis_group_scan: Σ bundle ranks %ld != group axis count %ld",
+             consumed_total, ngroup);
+  }
+  {
+    ca_size_t pv = 1;
+    for ( int bi = n_bundles - 1; bi >= 0; bi-- ) {
+      bundle_placeval[bi] = pv;
+      pv *= bundle_k[bi];
+    }
+  }
+
+  ca_size_t band = (group_prod > 0) ? (src->elements / group_prod) : 0;
+
+  /* --- supported dtype gate (CA_OBJECT handled by its own lane below) --- */
+  ca = src;
+  switch ( src->data_type ) {
+  case CA_BOOLEAN: case CA_INT8:  case CA_UINT8:  case CA_INT16: case CA_UINT16:
+  case CA_INT32:   case CA_UINT32: case CA_INT64: case CA_UINT64:
+  case CA_FLOAT32: case CA_FLOAT64: case CA_OBJECT: break;
+  default:
+    for ( int bi = 0; bi < n_bundles; bi++ ) ca_detach(bundle_ca[bi]);
+    rb_raise(rb_eRuntimeError,
+             "axis_group_scan: unsupported source data_type %d",
+             src->data_type);
+  }
+
+  /* --- output dtype per op: cumcount int64; cumsum / cumprod float64 (object
+     source -> object); cummax / cummin preserve the source dtype (object ->
+     object). --- */
+  int8_t out_dt;
+  if      ( op == GS_CUMCOUNT )                        { out_dt = CA_INT64; }
+  else if ( src->data_type == CA_OBJECT )              { out_dt = CA_OBJECT; }
+  else if ( op == GS_CUMSUM || op == GS_CUMPROD )      { out_dt = CA_FLOAT64; }
+  else                                                 { out_dt = src->data_type; }
+
+  /* --- source-shaped output --- */
+  ca_size_t odim[CA_RANK_MAX];
+  for ( int8_t i = 0; i < src->ndim; i++ ) odim[i] = src->dim[i];
+  VALUE vout = rb_carray_new(out_dt, src->ndim, odim, 0, NULL);
+  GetCArray(vout, co);
+
+  /* --- band raveled base address per band flat index (same construction as
+     the reduce min_addr / max_addr path): addr(cell) = band_addr[b] +
+     Σ sidx[s]*grstride[s].
+
+     CAREFUL: the walks below index the output by that raveled source address.
+     This is only valid because `vout` is a freshly allocated contiguous
+     entity of the source's shape, so output offset == source row-major
+     raveled address.  Handing a view here would scatter into wrong cells. --- */
+  ca_size_t rstride[CA_RANK_MAX], grstride[CA_RANK_MAX];
+  rstride[src->ndim - 1] = 1;
+  for ( int8_t k = (int8_t)(src->ndim - 2); k >= 0; k-- )
+    rstride[k] = rstride[k+1] * src->dim[k+1];
+  for ( long s = 0; s < ngroup; s++ ) grstride[s] = rstride[axes[s]];
+  int band_axis[CA_RANK_MAX]; int nband_axes = 0;
+  for ( int8_t k = 0; k < src->ndim; k++ )
+    if ( ! is_group[k] ) band_axis[nband_axes++] = k;
+  ca_size_t *band_addr = ALLOC_N(ca_size_t, band > 0 ? band : 1);
+  for ( ca_size_t bb = 0; bb < band; bb++ ) {
+    ca_size_t rem = bb, addr = 0;
+    for ( int j = nband_axes - 1; j >= 0; j-- ) {
+      ca_size_t d = src->dim[band_axis[j]];
+      addr += (rem % d) * rstride[band_axis[j]];
+      rem  /= d;
+    }
+    band_addr[bb] = addr;
+  }
+
+  /* band-independent scatter plan (built once on the first band, sized to the
+     slab = group_prod), + per-band dense [0, K_total) accumulators reset each
+     band.  accd (sum / prod), accn (running count) and seen (extremum first
+     member / object identity init) are all tiny; acco (object running VALUE) is
+     allocated only for the object arithmetic / extremum ops. */
+  ca_size_t  plan_n  = (group_prod > 0) ? group_prod : 1;
+  ca_size_t *sw_code = ALLOC_N(ca_size_t, plan_n);
+  ca_size_t *sw_doff = ALLOC_N(ca_size_t, plan_n);
+  ca_size_t *sw_moff = ALLOC_N(ca_size_t, plan_n);
+  ca_size_t *sw_addr = ALLOC_N(ca_size_t, plan_n);
+  double    *accd    = ALLOC_N(double,    K_total);
+  ca_size_t *accn    = ALLOC_N(ca_size_t, K_total);
+  char      *seen    = ALLOC_N(char,      K_total);
+  VALUE     *acco    = NULL;
+  double     acc_init = ( op == GS_CUMPROD ) ? 1.0 : 0.0;
+
+  boolean8_t *omask = NULL;
+  #define MARK_OUT_UNDEF(o) do {                   \
+      if ( ! omask ) {                             \
+        ca_create_mask(co);                        \
+        omask = (boolean8_t *) co->mask->ptr;      \
+      }                                            \
+      omask[o] = 1;                                \
+    } while (0)
+
+  if ( src->data_type == CA_OBJECT ) {
+    if ( op == GS_CUMCOUNT ) {                    /* int64 1-based running count */
+      int64_t *outi = (int64_t *) co->ptr;
+      GROUP_SCAN_OBJECT_WALK(
+        accn[code] += 1; outi[addr] = (int64_t) accn[code];,
+        outi[addr] = (int64_t) accn[code];
+      );
+    }
+    else {                                        /* object running VALUE */
+      VALUE *outo = (VALUE *) co->ptr;
+      acco = ALLOC_N(VALUE, K_total);
+      switch ( op ) {
+      case GS_CUMSUM:
+        /* Identity 0 (matching the core object cumsum): a group's acc is lazily
+           seeded to 0 (Fixnum, so 0 + ev promotes to ev's class), so a cell
+           masked before any present member holds 0 unmasked and an all-masked
+           group yields 0 everywhere.  seen doubles as the per-band init flag. */
+        GROUP_SCAN_OBJECT_WALK(
+          if ( ! seen[code] ) { acco[code] = INT2FIX(0); seen[code] = 1; }
+          acco[code] = rb_funcall(acco[code], rb_intern("+"), 1, ev);
+          outo[addr] = acco[code];,
+          if ( ! seen[code] ) { acco[code] = INT2FIX(0); seen[code] = 1; }
+          outo[addr] = acco[code];
+        );
+        break;
+      case GS_CUMPROD:
+        /* Identity 1, same lazy-init as cumsum. */
+        GROUP_SCAN_OBJECT_WALK(
+          if ( ! seen[code] ) { acco[code] = INT2FIX(1); seen[code] = 1; }
+          acco[code] = rb_funcall(acco[code], rb_intern("*"), 1, ev);
+          outo[addr] = acco[code];,
+          if ( ! seen[code] ) { acco[code] = INT2FIX(1); seen[code] = 1; }
+          outo[addr] = acco[code];
+        );
+        break;
+      case GS_CUMMAX:
+        /* No identity: seed from the first member, hold only once seen; a cell
+           masked before any member stays UNDEF (empty-max contract). */
+        GROUP_SCAN_OBJECT_WALK(
+          if ( ! seen[code] ) { acco[code] = ev; seen[code] = 1; }
+          else if ( NUM2INT(rb_funcall(ev, rb_intern("<=>"), 1, acco[code])) > 0 ) {
+            acco[code] = ev;
+          }
+          outo[addr] = acco[code];,
+          if ( seen[code] ) { outo[addr] = acco[code]; }
+          else { MARK_OUT_UNDEF(addr); }
+        );
+        break;
+      case GS_CUMMIN:
+        GROUP_SCAN_OBJECT_WALK(
+          if ( ! seen[code] ) { acco[code] = ev; seen[code] = 1; }
+          else if ( NUM2INT(rb_funcall(ev, rb_intern("<=>"), 1, acco[code])) < 0 ) {
+            acco[code] = ev;
+          }
+          outo[addr] = acco[code];,
+          if ( seen[code] ) { outo[addr] = acco[code]; }
+          else { MARK_OUT_UNDEF(addr); }
+        );
+        break;
+      }
+    }
+  }
+  else {                                          /* native dtype dispatch */
+    switch ( op ) {
+    case GS_CUMSUM: {
+      double *outd = (double *) co->ptr;
+      GROUP_SCAN_DISPATCH( accd[code] += v; outd[addr] = accd[code];,
+                           outd[addr] = accd[code]; );
+      break;
+    }
+    case GS_CUMPROD: {
+      double *outd = (double *) co->ptr;
+      GROUP_SCAN_DISPATCH( accd[code] *= v; outd[addr] = accd[code];,
+                           outd[addr] = accd[code]; );
+      break;
+    }
+    case GS_CUMCOUNT: {                           /* 1-based running count */
+      int64_t *outi = (int64_t *) co->ptr;
+      GROUP_SCAN_DISPATCH( accn[code] += 1; outi[addr] = (int64_t) accn[code];,
+                           outi[addr] = (int64_t) accn[code]; );
+      break;
+    }
+    case GS_CUMMAX:
+      GROUP_SCAN_EXTREMUM_DISPATCH( > );
+      break;
+    case GS_CUMMIN:
+      GROUP_SCAN_EXTREMUM_DISPATCH( < );
+      break;
+    }
+  }
+  #undef MARK_OUT_UNDEF
+
+  for ( int bi = 0; bi < n_bundles; bi++ ) ca_detach(bundle_ca[bi]);
+  xfree(band_addr);
+  xfree(sw_code);
+  xfree(sw_doff);
+  xfree(sw_moff);
+  xfree(sw_addr);
+  xfree(accd);
+  xfree(accn);
+  xfree(seen);
+  if ( acco ) xfree(acco);
+
+  RB_GC_GUARD(keep);
+  return vout;
+}
+
+void
+Init_ca_axis_group (void)
+{
+  rb_define_method(rb_cCArray, "__axis_group_reduce__",
+                   rb_ca_axis_group_reduce, 3);
+  rb_define_method(rb_cCArray, "__axis_group_scan__",
+                   rb_ca_axis_group_scan, 3);
+}