The naive sum_of_squares/count - mean^2 formula suffers catastrophic
cancellation for values with a large baseline and small spread (e.g.
values [1e8, 1e8+1] gave std 0.0 instead of 0.5).

- Distributed: switch the (n, 5) partial-stats layout from
  [count, sum, sum_of_squares, min, max] to the Welford/Chan
  representation [count, mean, M2, min, max]; merge via the Chan
  combine, finalize as variance = M2 / count.
- In-core complex features: compute the variance two-pass from the
  values already stored for the percentiles (exact).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

merge_block_edge_stats copied the entire (E, 5) global accumulator on
every call, making the documented per-block merge loop
O(number_of_blocks x global_edges). Mutate the caller's accumulator in
place instead (and return it, so existing loops keep working): one merge
is now O(block edges) regardless of global graph size. The Python
wrapper requires a C-contiguous writable float64 accumulator and no
longer coerces it (a silent copy would discard the update).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

The unordered_set added ~8x node/bucket overhead over the raw edge
data (about 114 MB extra for one million edges). Canonicalize into a
vector and sort + unique instead: same sorted-unique output, peak
memory close to one copy of the input. Document that outputs are valid
inputs, so callers can bound peak memory with a hierarchical merge.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

- own_begin / own_shape now require integer dtypes (floats such as
  [0.9, 0] were silently truncated by int()).
- Affinity offsets go through operator.index, which rejects floats.
- merge_edges rejects floating-dtype edge arrays and negative node ids
  in signed arrays (an int64 edge [-1, 2] silently wrapped to
  [2, 2**64 - 1] via the unconditional uint64 cast).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

The core statics return the raw _core.UndirectedGraph, so results
lacked the subclass convenience wrappers (find_edges on lists, ...) and
failed isinstance checks against bic.graph.UndirectedGraph. Add a
constructor overload taking (number_of_nodes, uvs, unique) — an
__init__ overload constructs the derived Python class, unlike a static
— and build the subclass classmethods on it. from_edges also gains the
single-pass C++ construction instead of insert_edges.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

The 2D/3D offset-box sweeps in feature_accumulation.hxx and the
owned-box sweeps in the distributed block extraction duplicated the
same clamp + loop logic. Move a single sweep_clipped_box_{2d,3d} with a
per-axis clip box into detail/grid.hxx; the in-core wrappers pass the
thread slab with unclipped remaining axes, the distributed dispatch
folds the owned box and slab into the clips. Inner loops are unchanged.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>