ci: green the minimal-install test lanes (skip optional-dep tests; resolve split json)

PR #211's macOS / minimal-install CI lane is red on pre-existing failures
unrelated to the malis change: several optional deps (numba, zarr, optuna,
em_erl, waterz) and the ABISS `ws` binary are absent there, and config_io
did not resolve per-split `json` paths.

- config_io.resolve_data_paths: resolve per-split `json` like image/label/mask
  (fixes test_resolve_data_paths_resolves_test_json). Also drops an unused
  inference_cfg local and wraps a long message (flake8 F841/E501) so the file
  stays lint-clean now that it is in the diff.
- guard optional-dependency tests with importorskip/skipif so they skip rather
  than error when the dep is absent, matching the existing zarr/tifffile idiom:
  waterz (both decoder test modules), optuna trial-timeout tests, zarr aux-cache
  test, em_erl nerl/skeleton tests, numba threshold-sensitivity, and the ABISS
  relative-script test (guarded on lib/abiss/build/ws).

Verified: under a simulated minimal install the previously-failing tests now
skip; with the deps installed they still pass.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: Donglai Wei

connectomics/config/pipeline/config_io.py

@@ -153,8 +153,7 @@ def _raise_unconsumed_keys(yaml_conf: DictConfig) -> None:
         "(sibling of `monitor`, `inference`, `decoding`, `tune`)."
     ),
     "use_timestamp": (
-        "field removed. Train mode is always timestamped; "
-        "test/tune modes are never timestamped."
+        "field removed. Train mode is always timestamped; " "test/tune modes are never timestamped."
     ),
 }
 _MONITOR_CHECKPOINT_ROOTS = (
@@ -201,8 +200,7 @@ def _reject_inference_runtime_alias_paths(explicit_field_paths: set[str]) -> Non
             alias_path = f"{root}.{alias}"
             if any(_path_is_or_descendant(path, alias_path) for path in explicit_field_paths):
                 raise ValueError(
-                    f"`{alias_path}` was renamed. "
-                    f"Use `{root}.{canonical_tail}` instead."
+                    f"`{alias_path}` was renamed. " f"Use `{root}.{canonical_tail}` instead."
                 )
 
     for root in _MONITOR_CHECKPOINT_ROOTS:
@@ -212,8 +210,7 @@ def _reject_inference_runtime_alias_paths(explicit_field_paths: set[str]) -> Non
                 if replacement.startswith("field "):
                     raise ValueError(f"`{alias_path}` {replacement}")
                 raise ValueError(
-                    f"`{alias_path}` was renamed. "
-                    f"Use `{root}.{replacement}` instead."
+                    f"`{alias_path}` was renamed. " f"Use `{root}.{replacement}` instead."
                 )
 
     # tune.output:* sub-block hoisted to tune.save_*
@@ -506,7 +503,6 @@ def validate_config(cfg: Config) -> None:
     if cfg.model.out_channels <= 0:
         raise ValueError("model.out_channels must be positive")
     model_heads = getattr(cfg.model, "heads", None) or {}
-    inference_cfg = getattr(cfg, "inference", None)
     inference_head = get_inference_model_value(cfg, "head", None)
     images_cfg = getattr(getattr(getattr(cfg, "monitor", None), "logging", None), "images", None)
     visualization_head = getattr(images_cfg, "head", None) if images_cfg is not None else None
@@ -556,8 +552,8 @@ def validate_config(cfg: Config) -> None:
             missing = [h for h in inference_head_names if h not in model_heads]
             if missing:
                 raise ValueError(
-                    f"inference.model.head={inference_head_names} references unknown heads {missing}; "
-                    f"available: {sorted(model_heads.keys())}."
+                    f"inference.model.head={inference_head_names} references unknown heads "
+                    f"{missing}; available: {sorted(model_heads.keys())}."
                 )
         if (
             visualization_head is not None
@@ -909,6 +905,11 @@ def _resolve_split_paths(split_cfg):
         split_cfg.image = _combine_path(split_base, split_cfg.image)
         split_cfg.label = _combine_path(split_base, split_cfg.label)
         split_cfg.mask = _combine_path(split_base, split_cfg.mask)
+        split_json_resolved = _combine_path(split_base, split_cfg.json)
+        if isinstance(split_json_resolved, list):
+            split_cfg.json = split_json_resolved[0] if split_json_resolved else None
+        else:
+            split_cfg.json = split_json_resolved
 
     # Resolve inference/test paths from merged runtime cfg.data.
     if getattr(cfg.data, "test", None) is not None:

tests/integration/test_affinity_cc3d.py

@@ -109,6 +109,7 @@ def test_basic_functionality(self, simple_affinities):
         ), f"Expected 3 labels (bg + 2 objects), got {len(unique_labels)}"
         assert 0 in unique_labels, "Background label 0 should be present"
 
+    @pytest.mark.skipif(not NUMBA_AVAILABLE, reason="numba required")
     def test_threshold_sensitivity(self, simple_affinities):
         """Test that threshold parameter affects segmentation."""
         # Low threshold - more connected

tests/test_decode_waterz.py

@@ -3,9 +3,15 @@
 from __future__ import annotations
 
 import numpy as np
+import pytest
 
 from connectomics.decoding.decoders import waterz as waterz_decoder
 
+# decode_waterz calls into the real waterz package (waterz._uint8,
+# merge_function_to_scoring); the fakes below only stub waterz.waterz(). Skip
+# the whole module when waterz is not installed (e.g. minimal CI).
+pytest.importorskip("waterz")
+
 
 class _FakeWaterzModule:
     """Minimal waterz stub for testing wrapper behavior."""

tests/unit/test_data_factory_zarr_aux.py

@@ -1,13 +1,15 @@
 from pathlib import Path
 
 import numpy as np
+import pytest
 
 from connectomics.config import Config
 from connectomics.data.io import save_volume
 from connectomics.training.lightning.data_factory import _maybe_precompute_label_aux
 
 
 def test_maybe_precompute_label_aux_reuses_existing_zarr_cache(monkeypatch, tmp_path: Path):
+    pytest.importorskip("zarr")
     label_path = tmp_path / "data.zarr" / "seg"
     aux_path = tmp_path / "data.zarr" / "seg_skeleton"
     save_volume(str(label_path), np.zeros((2, 2, 2), dtype=np.uint16), file_format="zarr")

tests/unit/test_decode_abiss_wrapper.py

@@ -2,13 +2,21 @@
 
 from __future__ import annotations
 
+import shutil
 import sys
+from pathlib import Path
 
 import numpy as np
 import pytest
 
 from connectomics.decoding import decode_abiss
 
+# The relative-script test below shells out to scripts/run_abiss_single.py, which
+# requires the compiled ABISS `ws` binary (lib/abiss/build/ws or on PATH). That
+# binary is absent in minimal installs (e.g. CI), so guard that test on it.
+_WS_BINARY = Path(__file__).resolve().parents[2] / "lib" / "abiss" / "build" / "ws"
+_ABISS_WS_AVAILABLE = _WS_BINARY.exists() or shutil.which("ws") is not None
+
 
 def test_decode_abiss_with_list_command_writes_npy_output():
     pred = np.zeros((3, 6, 8, 10), dtype=np.float32)
@@ -61,6 +69,7 @@ def test_decode_abiss_raises_if_output_missing():
         decode_abiss(pred, command=command)
 
 
+@pytest.mark.skipif(not _ABISS_WS_AVAILABLE, reason="ABISS ws binary not built (lib/abiss/build/ws)")
 def test_decode_abiss_with_relative_script_command_outside_repo_cwd(monkeypatch, tmp_path):
     pred = np.zeros((3, 6, 8, 10), dtype=np.float32)
     pred[:, 1:5, 2:7, 3:9] = 1.0

tests/unit/test_decode_waterz.py

@@ -7,6 +7,11 @@
 
 from connectomics.decoding.decoders import waterz as waterz_decoder
 
+# decode_waterz calls into the real waterz package (waterz._uint8,
+# merge_function_to_scoring); the fakes below only stub waterz.waterz(). Skip
+# the whole module when waterz is not installed (e.g. minimal CI).
+pytest.importorskip("waterz")
+
 
 class _FakeWaterzModule:
     """Minimal waterz stub for testing wrapper behavior."""

tests/unit/test_optuna_tuner.py

@@ -562,6 +562,7 @@ def _fake_decoder(predictions, **_kwargs):
 
 
 def test_objective_returns_bad_value_when_standard_trial_times_out(monkeypatch):
+    pytest.importorskip("optuna")
     cfg = Config()
     cfg.tune = TuneConfig()
     cfg.tune.trial_timeout = 12
@@ -588,6 +589,7 @@ def _raise_timeout(_evaluation_kind, _payload):
 
 
 def test_objective_returns_bad_value_when_waterz_batch_trial_times_out(monkeypatch):
+    pytest.importorskip("optuna")
     cfg = Config()
     cfg.tune = TuneConfig()
     cfg.tune.trial_timeout = 30

tests/unit/test_test_pipeline_multi_volume_eval.py

@@ -181,6 +181,7 @@ def test_run_evaluation_stage_accepts_plain_context_and_arrays():
 
 
 def test_compute_test_metrics_supports_nerl_without_dense_labels(tmp_path):
+    pytest.importorskip("em_erl")
     ERLGraph, _, _ = import_em_erl()
     graph_path = tmp_path / "gt_graph.npz"
     graph = ERLGraph(
@@ -233,6 +234,7 @@ def test_compute_test_metrics_supports_nerl_without_dense_labels(tmp_path):
 
 
 def test_compute_test_metrics_supports_banis_skeleton_pickle(tmp_path):
+    pytest.importorskip("em_erl")
     nx = pytest.importorskip("networkx")
     import pickle