fix: audit — score falsy-zero, crop person mismatch, adaptive cap, suppress_output, ldconfig glob, scheduler sleep

- immich_api: `score or confidence` treated 0.0 score as falsy; use explicit None check
- diversity: same falsy-zero fix in _get_face_confidence
- diversity: _crop_face_from_thumbnail scale loop now filters by person_id (was
  using first person's imageWidth/imageHeight regardless of target in group photos)
- jobs: partially-trained auto mode kept limit="auto" for adaptive stopping, then
  caps result to remaining capacity (was converting to int, silently disabling FPS
  adaptive threshold and early-stop)
- embeddings: _suppress_output finally block wraps first dup2 in try/finally so
  stderr is always restored even if stdout restore raises OSError
- Dockerfile: ldconfig find uses python3.* glob instead of hardcoded python3.13
- scheduler: sleep until next_run instead of fixed 60s; eliminates late-fire jitter
  and unnecessary wakeups on long schedules

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Dockerfile

@@ -75,7 +75,7 @@ COPY --from=build /usr/local/bin/uv /usr/local/bin/uv
 # Register every nvidia pip-package lib/ directory with ldconfig so that
 # onnxruntime-gpu and torch can find libcudnn, libcublas, libcufft, etc.
 # without a hand-maintained LD_LIBRARY_PATH. Skipped silently on cpu builds.
-RUN find /app/.venv/lib/python3.13/site-packages/nvidia -type d -name "lib" \
+RUN find /app/.venv/lib/python3.*/site-packages/nvidia -type d -name "lib" \
         2>/dev/null > /etc/ld.so.conf.d/nvidia-pip.conf && ldconfig || true
 
 RUN groupadd -g 568 apps && useradd -u 568 -g apps -m -s /bin/bash appuser \

scheduler.py

@@ -49,4 +49,4 @@ while True:
             logger.error(f"winnow run failed: {e}", exc_info=True)
             print(f"winnow run failed: {e}", flush=True)
         next_run = cron.get_next(float)
-    time.sleep(60)
+    time.sleep(max(1, next_run - time.time()))

winnow/diversity.py

@@ -105,7 +105,9 @@ def _get_face_confidence(asset: dict, person_id: str | None = None) -> float | N
             continue
         faces = person.get("faces", [])
         if faces:
-            return faces[0].get("score") or faces[0].get("confidence")
+            f = faces[0]
+            score = f.get("score")
+            return score if score is not None else f.get("confidence")
     return None
 
 
@@ -136,8 +138,10 @@ def _crop_face_from_thumbnail(
     x1, y1, x2, y2 = bbox
     img_w, img_h = img.size
 
-    # Get metadata dimensions to scale bbox
+    # Get metadata dimensions to scale bbox — must match the same person as _get_face_bbox
     for person in asset.get("people", []):
+        if person_id and person.get("id") != person_id:
+            continue
         faces = person.get("faces", [])
         if faces:
             meta_w = faces[0].get("imageWidth") or img_w

winnow/embeddings.py

@@ -33,8 +33,10 @@ def _suppress_output():
         os.dup2(devnull_fd, 2)
         yield
     finally:
-        os.dup2(saved_out, 1)
-        os.dup2(saved_err, 2)
+        try:
+            os.dup2(saved_out, 1)
+        finally:
+            os.dup2(saved_err, 2)
         os.close(devnull_fd)
         os.close(saved_out)
         os.close(saved_err)

winnow/immich_api.py

@@ -140,10 +140,11 @@ def fetch_face_data(asset_id: str, person_id: str | None = None) -> FaceData | N
             face.get("boundingBoxY2", 0),
         )
 
+        score = face.get("score")
         return FaceData(
             embedding=embedding,
             bbox=bbox,
-            confidence=face.get("score") or face.get("confidence"),
+            confidence=score if score is not None else face.get("confidence"),
             image_width=face.get("imageWidth", 0),
             image_height=face.get("imageHeight", 0),
         )

winnow/jobs.py

@@ -299,10 +299,13 @@ def auto_configure(people: list[dict]) -> list[dict]:
         has_embedding = is_embedding_available(entity_type)
         limit, selection_mode = _resolve_strategy(strategy, has_embedding)
 
-        # Cap selection to remaining capacity
+        # Cap selection to remaining capacity.
+        # For auto mode with partial training, keep "auto" so adaptive stopping
+        # still runs — just trim the result to the remaining capacity afterward.
+        auto_cap = None
         if limit == "auto":
             if already_uploaded > 0:
-                limit = capacity  # partially filled — select exactly what remains
+                auto_cap = capacity
         else:
             limit = min(limit, capacity)
 
@@ -312,6 +315,8 @@ def auto_configure(people: list[dict]) -> list[dict]:
         selected_assets = _perform_selection(
             recent_assets, limit, name, selection_mode, entity_type, person_id=person["id"]
         )
+        if auto_cap is not None:
+            selected_assets = selected_assets[:auto_cap]
 
         if selected_assets:
             rprint(f"  [green]Queued {len(selected_assets)} images for {name}.[/green]")