Reorganize tutorials, remove augmentation preset, fix skeleton coordinate bug

- Move mitoEM configs to tutorials/mitoEM/ with shortened names (H, R, HR, common)
- Move neuron_nisb configs to tutorials/neuron_nisb/ subdirectory
- Move config profiles/templates from configs/ to connectomics/config/
- Remove augmentation.preset tri-state; use per-transform enabled flags directly
- Fix _batch_skeletonize: divide kimimaro physical coords by resolution to get voxel indices
- Add root_path to mito_betaseg.yaml to simplify train/val/test paths
- Add test_sdt_precomputed.py: verify precomputed skeleton SDT matches direct computation
- Clean up obsolete tutorials/misc/ files and test_hydra_config.py

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

Author: Donglai Wei

.claude/reference/abiss.md

@@ -0,0 +1,150 @@
+# em_pipeline Reference
+
+**Location**: `/projects/weilab/weidf/lib/em_pipeline`
+
+Multi-GPU/multi-core pipeline engine for large-scale 3D EM neuron segmentation and reconstruction. Implements a two-stage approach: voxel→supervoxel (waterz watershed) then supervoxel→instance (region graph + branch resolution). Designed for distributed, chunked processing of datasets like zebrafish brain (5700x10913x10664 voxels).
+
+## Directory Structure
+
+```
+em_pipeline/
+├── main.py                      # CLI entry point (106 lines), SLURM integration
+├── test.py                      # Dev/visualization script (983 lines)
+├── conf/
+│   ├── j0126.yml                # Project config (zebrafish example)
+│   └── cluster.yml              # Cluster/environment config
+├── db/                          # Local database/cache (HDF5 files)
+├── em_pipeline/
+│   ├── tasks/
+│   │   ├── __init__.py          # Task factory
+│   │   ├── task.py              # Base Task class (34 lines)
+│   │   ├── waterz.py            # Waterz segmentation tasks (217 lines)
+│   │   ├── branch.py            # Branch resolution tasks (363 lines)
+│   │   ├── region_graph.py      # Region graph / soma BFS (104 lines)
+│   │   └── eval.py              # ERL evaluation (25 lines)
+│   └── lib/
+│       └── rpca.py              # Robust PCA (68 lines)
+├── setup.py
+├── environment.yml
+└── README.md
+```
+
+~1,933 lines total.
+
+## Architecture
+
+### Two-Stage Pipeline
+
+```
+Affinity Predictions (HDF5/Zarr)
+  ↓  Stage 1: Voxel → Supervoxel
+Waterz Segmentation (per-chunk supervoxels)
+  ↓  Chunk merging (global segment IDs)
+  ↓  Soma constraint (BFS-based soma assignment)
+  ↓  Stage 2: Supervoxel → Instance
+Branch Resolution (S1/S2/S3 IOU+affinity checks)
+  ↓
+Skeleton Generation (multi-scale morphology)
+```
+
+### Task Class Hierarchy
+
+```
+Task (base)
+├── WaterzTask           # Chunk-wise waterz segmentation
+├── WaterzSoma2DTask     # 2D soma-aware waterz (per z-slice)
+├── WaterzStatsTask      # Consolidate stats across chunks
+├── BranchChunkTask      # Per-chunk branch resolution
+│   ├── BranchBorderTask # Cross-chunk boundary handling
+│   └── BranchAllTask    # Global aggregation + skeletonization
+├── RegionGraphChunkTask # Soma constraints on region graphs
+├── RegionGraphBorderTask
+└── ERLTask              # Evaluation against ground truth
+```
+
+## Usage
+
+```bash
+# CLI entry point
+python main.py -c conf/j0126.yaml -t [task] -i [job_id] -n [job_num] -nc [num_cpu]
+
+# Stage 1: Voxel → Supervoxel
+python main.py -c conf/j0126.yaml -t waterz
+python main.py -c conf/j0126.yaml -t waterz-stats
+python main.py -c conf/j0126.yaml -t rg-border
+python main.py -c conf/j0126.yaml -t rg-all
+
+# Stage 2: Supervoxel → Instance
+python main.py -c conf/j0126.yaml -t branch-border -o relabel
+python main.py -c conf/j0126.yaml -t branch-all -o s2-4-8-8
+```
+
+CLI arguments: `-c` config file, `-t` task name, `-i` job index, `-n` total jobs, `-nc` CPUs per job, `-p` SLURM partition.
+
+## Configuration (YAML)
+
+### Project Config (`conf/j0126.yml`)
+
+| Section | Keys | Purpose |
+|---------|------|---------|
+| `im` | path, shape, tile_shape, res | Input image volume |
+| `mask` | blood_vessel, soma, border, soma_ratio, soma_id0/id1 | Segmentation constraints |
+| `aff` | path, aff_shape, low, high | Affinity predictions + thresholds |
+| `waterz` | mf, thres, num_z, nb, opt_frag, small_size, small_aff, small_dust, bg_thres | Watershed parameters |
+| `branch` | s1_iou, s1_sz, s1_rg, s3_iou, s3_sz, skel_dust | Branch resolution thresholds |
+| `rg` | thres_z | Region graph parameters |
+| `output` | path | Output directory |
+| `eval` | val, test | Evaluation datasets |
+
+### Cluster Config (`conf/cluster.yml`)
+
+Keys: `folder`, `env` (setup commands), `python` (executable), `num_gpu`, `memory`.
+
+## Key Algorithms
+
+### Waterz Task
+- Applies affinity masks (blood vessel, border)
+- Runs waterz agglomeration with configurable merge function and threshold
+- Generates per-chunk region graphs
+- Output: HDF5 with `seg`, `id`, `score` datasets
+
+### Soma-Aware Waterz (2D)
+- Processes each z-slice independently
+- Integrates soma mask constraints via seeded watershed
+- Resolves soma-based false splits/merges
+
+### Branch Resolution (3 stages)
+- **S1**: IOU-based merge scoring within chunks
+- **S2**: IOU best-buddy pairing (bidirectional agreement)
+- **S3**: One-sided IOU + affinity validation for remaining candidates
+
+### Soma BFS
+- Breadth-first search to grow soma regions outward
+- Assigns non-soma segments to nearest soma
+- Handles ambiguous segments (multiple soma connections)
+
+## Data Formats
+
+- **HDF5** (.h5): Primary I/O for volumes and results
+- **Zarr** (.zarr): Large-scale chunked arrays
+- **Pickle**: Dask arrays and serialized objects
+- **PNG/TIFF**: 2D mask/image files
+
+## Dependencies
+
+### Core (from environment.yml)
+numpy 1.24.3, scipy, h5py, pyyaml 6.0.2, pillow 10.4.0, cc3d 3.18.0, fastremap 1.15.0, mahotas 1.4.18, zarr 2.16.1, dask 2023.5.0, networkx 3.1, cloudpickle 3.0.0
+
+### External (installed separately)
+- **em_util**: `git clone git@github.com:PytorchConnectomics/em_util.git && pip install -e .`
+- **waterz**: `git clone -b affuint8 git@github.com:donglaiw/waterz.git && pip install -e .`
+- **zwatershed**: `git clone git@github.com:donglaiw/zwatershed.git && pip install -e .`
+- **em_erl**: Evaluation utilities (referenced in code)
+
+## Design Principles
+
+1. **Chunked processing**: Overlapping chunks for parallel execution on large volumes
+2. **Hierarchical resolution**: Sequential refinement from coarse to fine
+3. **Constraint integration**: Soma masks and boundary constraints guide segmentation
+4. **IOU-based merging**: Intersection-over-union drives segment agglomeration
+5. **Skeleton output**: Multi-scale neuron skeletons for morphological analysis