fix(loader): authenticate h5ad loads in dataset metadata tools

dataset_count_metadata_values / dataset_crosstab_metadata_values /
dataset_validate_manifest_against_data treated adata_path as a local path
(Path(url).exists() -> False), so a private pdac-research-data h5ad URL wrongly
returned 'File not found' and never authenticated. These three were missed in
the loader-auth centralization (ce5168d/04b26ba); route them through
src.core.data_io.resolve_to_local_path (HF_TOKEN auth for private
huggingface.co /resolve/ URLs, temp-file cleanup) like the sibling tools.

Surfaced live: the XDI-002 late cross-dataset eval stalled when the agent could
not load paca_au_rnaseq via dataset_count_metadata_values and fell back to an
unauthenticated urllib.urlretrieve that hung on the private repo. Verified the
fixed tool now loads the private paca_au_rnaseq.h5ad and returns real Bailey
subtype counts.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

src/tools/dataset_tools/metadata.py

@@ -203,15 +203,30 @@ def dataset_count_metadata_values(
     import scanpy as sc
     from pathlib import Path
     from collections import Counter as _Counter
+    from src.core.data_io import resolve_to_local_path
 
-    data_file = Path(adata_path)
-    if not data_file.exists():
+    # Resolve local paths AND remote URLs (e.g. private pdac-research-data h5ads)
+    # to a readable local file, authenticating via HF_TOKEN for private
+    # huggingface.co /resolve/ URLs. Without this, an https URL was treated as a
+    # local path (Path.exists() == False) and wrongly reported "File not found".
+    try:
+        local_path, is_temp = resolve_to_local_path(adata_path)
+    except Exception as e:
+        return {"error": f"Could not resolve {adata_path}: {e}"}
+
+    if not is_temp and not Path(local_path).exists():
         return {"error": f"File not found: {adata_path}"}
 
     try:
-        adata = sc.read_h5ad(adata_path)
+        adata = sc.read_h5ad(local_path)
     except Exception as e:
         return {"error": f"Could not load h5ad: {e}"}
+    finally:
+        if is_temp:
+            try:
+                Path(local_path).unlink()
+            except OSError:
+                pass
 
     try:
         manifest = load_manifest(dataset_id)
@@ -335,15 +350,27 @@ def dataset_crosstab_metadata_values(
     import scanpy as sc
     import pandas as pd
     from pathlib import Path
+    from src.core.data_io import resolve_to_local_path
 
-    data_file = Path(adata_path)
-    if not data_file.exists():
+    # Resolve local paths AND remote URLs (incl. private HF repos via HF_TOKEN).
+    try:
+        local_path, is_temp = resolve_to_local_path(adata_path)
+    except Exception as e:
+        return {"error": f"Could not resolve {adata_path}: {e}"}
+
+    if not is_temp and not Path(local_path).exists():
         return {"error": f"File not found: {adata_path}"}
 
     try:
-        adata = sc.read_h5ad(adata_path)
+        adata = sc.read_h5ad(local_path)
     except Exception as e:
         return {"error": f"Could not load h5ad: {e}"}
+    finally:
+        if is_temp:
+            try:
+                Path(local_path).unlink()
+            except OSError:
+                pass
 
     try:
         manifest = load_manifest(dataset_id)
@@ -442,11 +469,25 @@ def dataset_validate_manifest_against_data(
     import scanpy as sc
 
     from src.workflows.manifest_data_validation import validate_manifest_against_data
+    from src.core.data_io import resolve_to_local_path
+    from pathlib import Path as _Path
+
+    # Resolve local paths AND remote URLs (incl. private HF repos via HF_TOKEN).
+    try:
+        local_path, _is_temp = resolve_to_local_path(adata_path)
+    except Exception as e:
+        return {"error": f"Could not resolve {adata_path}: {e}"}
 
     try:
-        adata = sc.read_h5ad(adata_path)
+        adata = sc.read_h5ad(local_path)
     except Exception as e:
         return {"error": f"Could not load h5ad at '{adata_path}': {e}"}
+    finally:
+        if _is_temp:
+            try:
+                _Path(local_path).unlink()
+            except OSError:
+                pass
 
     try:
         manifest = load_manifest(dataset_id)