Add waterz region graph + merge_id, branch_merge with singleton, IoU module, oracle ARE, bbox cleanup

waterz (lib/waterz):
- Add get_region_graph() using JIT scoring functions (any merge_function)
  with channel masking (all/z/xy) via affinity zeroing
- Add merge_id() C++/Cython: union-find merge on edge lists (4 modes:
  ID-only, +affinity, +count, +both), ported from waterz_dw
- Clean up: remove standalone C++ region graph scan, single get_region_graph
  entry point for all scoring functions
- frontend_merge.cpp now contains only merge_segments

connectomics:
- Add data/process/iou.py: bbox-accelerated seg_to_iou and segs_to_iou
  ported from em_util
- Add data/process/bbox.py: replace find_objects with em_util row/col scan
- Add metrics/segmentation_numpy.py: adapted_rand_oracle for incremental
  per-GT oracle ARE (O(nnz_per_row) instead of O(volume))
- Add decoding/decoders/branch_merge.py: singleton_size_ratio parameter
  for merging single-slice fragments via size ratio
- Update decode_waterz: dust_merge_size/dust_merge_affinity/dust_remove_size
  renamed parameters

Reference docs: em_util.md, waterz_dw.md updated

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: Donglai Wei

.claude/reference/em_util.md

@@ -0,0 +1,95 @@
+# em_util Reference
+
+**Location:** `/projects/weilab/weidf/lib/em_util`
+**GitHub:** https://github.com/PytorchConnectomics/em_util
+**License:** MIT
+
+Utility library for EM connectomics: volume I/O, segmentation operations, evaluation metrics, neuroglancer visualization, WebKnossos integration, and SLURM job management.
+
+## Module Overview
+
+| Module | Purpose |
+|--------|---------|
+| `em_util.io` | Universal I/O (HDF5, TIFF, PNG, Zarr, CloudVolume), bbox, skeleton, chunked ops, UnionFind |
+| `em_util.seg` | Segmentation ops: relabel, remove small, morphology, connected components, IoU |
+| `em_util.eval` | Metrics: adapted_rand, VOI, confusion matrix |
+| `em_util.ng` | Neuroglancer 3D visualization |
+| `em_util.wk` | WebKnossos remote dataset management |
+| `em_util.vast` | VAST annotation format support |
+| `em_util.cluster` | SLURM job submission utilities |
+
+## Conventions
+
+- **Volume axis order:** ZYX (depth, height, width)
+- **Bounding boxes:** 2D: `[seg_id, y0, y1, x0, x1, count]`, 3D: `[seg_id, z0, z1, y0, y1, x0, x1, count]`
+- **Segmentation:** Integer arrays, 0 = background
+
+## Key Functions
+
+### I/O (`em_util.io`)
+
+```python
+read_vol(filename, dataset=None)          # Universal reader (.h5, .tif, .npy, .zarr, etc.)
+read_h5(filename, dataset=None)           # HDF5 reader
+write_h5(filename, data, dataset="main")  # HDF5 writer (gzip compressed)
+read_image(filename, image_type="image")  # Single image reader
+read_image_folder(filename, ...)          # Image stack reader (glob patterns)
+compute_bbox_all(seg, do_count=False)     # Bounding boxes for all segments
+vol_to_skel(labels, res=(32,32,30))       # Kimimaro skeletonization
+```
+
+### Segmentation (`em_util.seg`)
+
+```python
+seg_to_count(seg, do_sort=True)           # Segment sizes (sorted)
+seg_relabel(seg, do_sort=True)            # Relabel by size (largest=1)
+seg_remove_small(seg, threshold=100)      # Remove small segments
+seg_to_cc(seg)                            # Connected component relabeling
+seg_biggest_cc(seg)                       # Keep only largest CC per label
+```
+
+### IoU (`em_util.seg.iou`) — Key for branch merge
+
+```python
+seg_to_iou(seg0, seg1, uid0=None, bb0=None, uid1=None, uc1=None, th_iou=0)
+```
+Compute IoU between two segmentations (2D or 3D). Uses bounding-box-accelerated overlap counting — only scans the bbox region of each segment in seg0, masking against seg1.
+
+Returns `(N, 5)` array: `[seg_id, best_match_id, count0, count1, overlap_count]`
+
+```python
+segs_to_iou(get_seg, index, th_iou=0)
+```
+Track segments across z-slices. `get_seg(i)` returns 2D segmentation at slice `i`. Iterates consecutive pairs, computing IoU with bbox acceleration. Returns list of overlap matrices (one per boundary).
+
+**Performance:** Bbox-accelerated — only scans pixels within each segment's bounding box, making it fast for sparse segmentations where segments occupy a small fraction of the image.
+
+### Evaluation (`em_util.eval`)
+
+```python
+adapted_rand(seg, gt, all_stats=False)    # SNEMI3D adapted rand error
+voi(seg, gt)                              # VOI (split, merge)
+```
+
+### Chunked Processing (large volumes)
+
+```python
+vol_func_chunk(input_file, vol_func, ...)        # Apply function chunk-by-chunk
+vol_downsample_chunk(input_file, ratio=[1,2,2])  # Downsample large HDF5
+compute_bbox_all_chunk(seg_file, chunk_num=1)     # Bbox from large file
+```
+
+### UnionFind (`em_util.io`)
+
+```python
+uf = UnionFind(elements)
+uf.union(a, b)                 # Merge two sets
+uf.find(a)                     # Find root
+uf.components()                # List all components
+uf.component_relabel_arr()     # Numpy relabel array
+```
+
+## Dependencies
+
+**Core:** numpy, scipy, h5py, imageio, scikit-image, tqdm, cc3d, pyyaml, networkx
+**Optional:** cloudvolume, zarr, kimimaro, neuroglancer, webknossos, igneous