fix/embedding unique obs names

applications/dynaclr/scripts/pseudotime/0-select_candidates/manual_candidates.py

@@ -117,7 +117,6 @@ def _validate_against_anndata(dataset_id: str, fov_name: str, tracks: list[dict]
     ds_meta = DATASETS[dataset_id]
     emb_path = ds_meta["embedding_zarr"]
     adata = ad.read_zarr(emb_path)
-    adata.obs_names_make_unique()
     obs = adata.obs
     fov_obs = obs[obs["fov_name"].astype(str) == fov_name]
 

applications/dynaclr/scripts/pseudotime/0-select_candidates/select_candidates.py

@@ -195,7 +195,6 @@ def _select_productive_from_zarr(
         return pd.DataFrame(columns=OUTPUT_COLUMNS)
 
     adata = ad.read_zarr(matches[0])
-    adata.obs_names_make_unique()
     if pred_column not in adata.obs.columns:
         _logger.warning(f"[{dataset_id}] {pred_column} not in {matches[0].name}; productive empty")
         return pd.DataFrame(columns=OUTPUT_COLUMNS)
@@ -296,7 +295,6 @@ def _select_mock_from_zarr(
         return pd.DataFrame(columns=OUTPUT_COLUMNS)
 
     adata = ad.read_zarr(matches[0])
-    adata.obs_names_make_unique()
     obs = adata.obs.copy()
     obs = obs[obs["fov_name"].astype(str).str.contains(fov_pattern, regex=False)]
     if obs.empty:
@@ -413,7 +411,6 @@ def _load_lc_predictions(
         )
 
     adata = ad.read_zarr(matches[0])
-    adata.obs_names_make_unique()
     if pred_column not in adata.obs.columns:
         _logger.warning(f"{pred_column} not in {matches[0]} .obs; LC fallback")
         return pd.DataFrame()

applications/dynaclr/scripts/pseudotime/2-align_cells/align_embedding.py

@@ -240,7 +240,6 @@ def _load_query_embeddings(
         zarr_path = find_embedding_zarr(ds_cfg["pred_dir"], prefix + embedding_pattern)
 
         adata = ad.read_zarr(zarr_path)
-        adata.obs_names_make_unique()
 
         # FOV restriction from the dataset config (e.g. "C/2") — keeps us
         # out of control wells unless the user explicitly wants them.

applications/dynaclr/scripts/pseudotime/2-align_cells/align_lc.py

@@ -101,7 +101,6 @@ def _load_lc_predictions(
             f"{[m.name for m in matches]}; using first"
         )
     adata = ad.read_zarr(matches[0])
-    adata.obs_names_make_unique()
     if pred_column not in adata.obs.columns:
         _logger.warning(f"{pred_column} not in {matches[0]} .obs")
         return pd.DataFrame(columns=["fov_name", "track_id", "t", pred_column])

applications/dynaclr/scripts/pseudotime/3-organelle-remodeling/readout_common.py

@@ -89,7 +89,6 @@ def load_organelle_embeddings(
             _logger.warning(f"[{ds_id}] no embedding zarr matched {prefix + embedding_pattern}: {exc}")
             continue
         adata = ad.read_zarr(zarr_path)
-        adata.obs_names_make_unique()
         out[ds_id] = adata
         _logger.info(f"[{ds_id}] loaded {Path(zarr_path).name} ({adata.n_obs} cells)")
     return out

applications/dynaclr/src/dynaclr/evaluation/linear_classifiers/orchestrated.py

@@ -73,7 +73,6 @@ def run_linear_classifiers(
             raise FileNotFoundError(f"No .zarr files found in {embeddings_path}")
         parts = [ad.read_zarr(p) for p in zarr_paths]
         adata = ad.concat(parts, join="outer")
-        adata.obs_names_make_unique()
         click.echo(f"  Loaded {len(zarr_paths)} per-experiment zarrs")
     else:
         adata = ad.read_zarr(embeddings_path)

applications/dynaclr/src/dynaclr/evaluation/mmd/compute_mmd.py

@@ -581,7 +581,6 @@ def run_mmd_pooled(config: MMDPooledConfig) -> pd.DataFrame:
 
     adatas = [ad.read_zarr(p) for p in config.input_paths]
     combined = ad.concat(adatas, join="outer", label="source_experiment")
-    combined.obs_names_make_unique()
 
     if config.obs_filter:
         mask = pd.Series([True] * len(combined), index=combined.obs.index)

packages/viscy-utils/src/viscy_utils/callbacks/embedding_writer.py

@@ -1,6 +1,7 @@
 """Callback for writing embeddings to zarr store."""
 
 import logging
+import uuid
 from pathlib import Path
 from typing import Any, Dict, Literal, Optional, Sequence
 
@@ -166,6 +167,11 @@ def write_embedding_dataset(
         elif hasattr(s, "cat") and isinstance(s.cat.categories.dtype, pd.StringDtype):
             ultrack_indices[col] = s.cat.rename_categories(s.cat.categories.astype(object))
 
+    # obs_names are an opaque, unique handle — never referenced directly (cell identity
+    # lives in the obs columns). Random UUIDs stay unique within a store and across any
+    # number of concatenated stores, so consumers never need obs_names_make_unique().
+    ultrack_indices.index = [str(uuid.uuid4()) for _ in range(len(ultrack_indices))]
+
     if embedding_key == "projections":
         if projections is None:
             raise ValueError("embedding_key='projections' requires projections to be provided.")
@@ -273,7 +279,7 @@ def write_on_epoch_end(
         """Write predictions and dimensionality reductions to a zarr store."""
         features = _move_and_stack_embeddings(predictions, "features")
         projections = _move_and_stack_embeddings(predictions, "projections")
-        ultrack_indices = pd.concat([pd.DataFrame(p["index"]) for p in predictions])
+        ultrack_indices = pd.concat([pd.DataFrame(p["index"]) for p in predictions], ignore_index=True)
 
         write_embedding_dataset(
             output_path=self.output_path,

packages/viscy-utils/tests/test_embedding_writer.py

@@ -0,0 +1,45 @@
+import anndata as ad
+import numpy as np
+import pandas as pd
+
+from viscy_utils.callbacks.embedding_writer import write_embedding_dataset
+
+
+def _make_index_df(n: int, fov: str) -> pd.DataFrame:
+    # Mimic EmbeddingWriter.write_on_epoch_end: per-batch frames concatenated, so the
+    # positional index restarts each batch and the result has duplicate labels.
+    half = n // 2
+    batches = [
+        pd.DataFrame({"fov_name": [fov] * half, "track_id": range(half), "t": range(half)}),
+        pd.DataFrame(
+            {"fov_name": [fov] * (n - half), "track_id": range(n - half), "t": range(n - half)}
+        ),
+    ]
+    df = pd.concat(batches)
+    assert df.index.has_duplicates  # precondition: the input that used to leak into obs_names
+    return df
+
+
+def test_obs_names_unique_within_store(tmp_path):
+    n, d = 6, 4
+    features = np.random.default_rng(0).standard_normal((n, d)).astype(np.float32)
+    out = tmp_path / "store.zarr"
+
+    write_embedding_dataset(output_path=out, features=features, index_df=_make_index_df(n, "A/1"))
+
+    adata = ad.read_zarr(out)
+    assert adata.obs_names.is_unique
+
+
+def test_obs_names_unique_across_concatenated_stores(tmp_path):
+    rng = np.random.default_rng(0)
+    n, d = 6, 4
+    paths = []
+    for i, fov in enumerate(["A/1", "B/2"]):
+        out = tmp_path / f"store_{i}.zarr"
+        features = rng.standard_normal((n, d)).astype(np.float32)
+        write_embedding_dataset(output_path=out, features=features, index_df=_make_index_df(n, fov))
+        paths.append(out)
+
+    combined = ad.concat([ad.read_zarr(p) for p in paths])
+    assert combined.obs_names.is_unique