fix: sync CAMS review fixes from upstream PR branch

Cherry-pick atomic download, tz-aware available(), and coordinate-based
lead-time selection fixes from feat/cams-datasource (NVIDIA#780).

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

Author: claude

CLAUDE.md

@@ -3,13 +3,15 @@
 ## Repository
 
 Fork von [NVIDIA/earth2studio](https://github.com/NVIDIA/earth2studio). Upstream-Sync via `git fetch upstream && git merge upstream/main`.
+Upstream-Version: 0.14.0a0 (gemergt 2026-03-29). Upstream-PR: NVIDIA/earth2studio#780 (CAMS DataSource).
 
 **Fork-Strategie:** Upstream-Files NICHT modifizieren. Eigene Additions in separaten Files halten. Nur `__init__.py` Imports und `pyproject.toml` Extras minimal patchen.
 
 Eigene Erweiterungen:
 - CAMS DataSource/Lexicon (`earth2studio/data/cams.py`, `earth2studio/lexicon/cams.py`)
 - Serve Workflows (`serve/server/example_workflows/`)
 - Client SDK (`serve/client/earth2studio_client/`)
+- Remote Examples (`examples/remote/`)
 - Windows Build Scripts (`scripts/`)
 
 ## Project Rules
@@ -74,6 +76,9 @@ Config: `serve/server/conf/config.yaml`. Custom Workflows via `WORKFLOW_DIR` env
 - **Zarr IO erwartet Tensors**: `io.write()` braucht `torch.Tensor`, nicht numpy arrays.
 - **PyTorch SHMEM**: Container braucht `ipc: host` und `ulimits` (memlock, stack).
 - **fetch_data -> map_coords**: `fetch_data()` liefert GFS-Koordinaten (721 lat), aber FCN erwartet 720. Bei custom Workflows nach `fetch_data` immer `map_coords(x, coords, model.input_coords())` aufrufen.
+- **pyproject.toml workspace**: `earth2studio-client` braucht SOWOHL `[tool.uv.workspace] members` ALS AUCH `{ workspace = true }` in `[tool.uv.sources]`.
+- **Examples-Struktur**: Upstream nutzt Unterordner (`examples/01_getting_started/`, etc.). Eigene Examples NUR in `examples/remote/`.
+- **Upstream-PR Branch**: Immer von `upstream/main` aus erstellen, nur eigene Files cherry-picken.
 
 ## Client-SDK
 

earth2studio/data/cams.py

@@ -16,8 +16,10 @@
 
 import asyncio
 import hashlib
+import os
 import pathlib
 import shutil
+import tempfile
 from dataclasses import dataclass
 from datetime import datetime, timedelta
 from time import sleep
@@ -104,12 +106,22 @@ def _download_cams_netcdf(
             )
         else:
             sleep(2.0)
-    r.download(str(cache_path))
+    tmp_fd, tmp_name = tempfile.mkstemp(dir=cache_path.parent, suffix=".nc.tmp")
+    try:
+        os.close(tmp_fd)
+        r.download(tmp_name)
+        os.replace(tmp_name, cache_path)
+    except Exception:
+        pathlib.Path(tmp_name).unlink(missing_ok=True)
+        raise
     return cache_path
 
 
 def _extract_field(
-    ds: xr.Dataset, nc_key: str, level: str = "0", time_index: int = 0
+    ds: xr.Dataset,
+    nc_key: str,
+    level: str = "0",
+    lead_time_hours: int | None = None,
 ) -> np.ndarray:
     if nc_key not in ds:
         raise ValueError(
@@ -118,21 +130,22 @@ def _extract_field(
         )
     field = ds[nc_key]
     non_spatial = [d for d in field.dims if d not in ("latitude", "longitude")]
-    sel: dict[str, int] = {}
+    isel: dict[str, int] = {}
     for d in non_spatial:
         if d == "level" and level:
             level_val = float(level) if level else 0.0
             level_coords = field.coords["level"].values
             nearest_idx = int(np.argmin(np.abs(level_coords - level_val)))
-            sel[d] = nearest_idx
-        elif d in ("time", "forecast_period", "forecast_reference_time"):
-            sel[d] = time_index
-        elif field.sizes[d] > 1:
-            sel[d] = 0
+            isel[d] = nearest_idx
+        elif d == "forecast_period" and lead_time_hours is not None:
+            fp_vals = field.coords["forecast_period"].values.astype(float)
+            target = float(lead_time_hours)
+            nearest_idx = int(np.argmin(np.abs(fp_vals - target)))
+            isel[d] = nearest_idx
         else:
-            sel[d] = 0
-    if sel:
-        field = field.isel(sel)
+            isel[d] = 0
+    if isel:
+        field = field.isel(isel)
     return field.values
 
 
@@ -288,7 +301,10 @@ def available(
         """
         if isinstance(time, datetime):
             time = [time]
-        return all(t >= _CAMS_EU_MIN_TIME for t in time)
+        return all(
+            (t.replace(tzinfo=None) if t.tzinfo else t) >= _CAMS_EU_MIN_TIME
+            for t in time
+        )
 
     @staticmethod
     def _validate_time(times: list[datetime]) -> None:
@@ -509,7 +525,10 @@ def available(
         """
         if isinstance(time, datetime):
             time = [time]
-        return all(t >= _CAMS_GLOBAL_MIN_TIME for t in time)
+        return all(
+            (t.replace(tzinfo=None) if t.tzinfo else t) >= _CAMS_GLOBAL_MIN_TIME
+            for t in time
+        )
 
     @staticmethod
     def _validate_time(times: list[datetime]) -> None:
@@ -568,12 +587,13 @@ def _fetch_forecast(
             },
         )
 
-        for lt_idx in range(len(lead_hours)):
+        for lt_idx, lt_h in enumerate(lead_hours):
             for info in var_infos:
                 _, modifier = CAMSLexicon[info.e2s_name]
                 da[0, lt_idx, info.index] = modifier(
                     _extract_field(
-                        ds, info.nc_key, level=info.level, time_index=lt_idx
+                        ds, info.nc_key, level=info.level,
+                        lead_time_hours=int(lt_h),
                     )
                 )
 

test/data/test_cams.py

@@ -96,6 +96,8 @@ def test_cams_time_validation():
 def test_cams_available():
     assert CAMS.available(datetime.datetime(2024, 1, 1))
     assert not CAMS.available(datetime.datetime(2015, 1, 1))
+    # timezone-aware datetimes must not raise TypeError
+    assert CAMS.available(datetime.datetime(2024, 1, 1, tzinfo=datetime.UTC))
 
 
 # ---- CAMS_FX tests ----
@@ -134,3 +136,5 @@ def test_cams_fx_fetch(variable, lead_time):
 def test_cams_fx_available():
     assert CAMS_FX.available(datetime.datetime(2024, 1, 1))
     assert not CAMS_FX.available(datetime.datetime(2010, 1, 1))
+    # timezone-aware datetimes must not raise TypeError
+    assert CAMS_FX.available(datetime.datetime(2024, 1, 1, tzinfo=datetime.UTC))