Fix the bugs reported against 0.19.0

app/core/config.py

@@ -28,11 +28,23 @@ def _env_path_opt(name: str) -> Path | None:
     return Path(raw).expanduser().resolve() if raw else None
 
 
+# Why torch could not be loaded at the last probe, or None if it loaded (#730).
+# Refreshed on every probe rather than decided once, so a repaired install
+# clears it without a restart.
+_torch_load_error: str | None = None
+
+
+def torch_load_error() -> str | None:
+    """The last probe's reason torch is installed but unusable, if any."""
+    return _torch_load_error
+
+
 def available_torch_devices() -> list[str]:
     """Compute devices this machine can actually use, best-first. CPU is always
     present; cuda/mps depend on the hardware + installed torch build. The
     Settings UI uses this to disable options that aren't available/detected so
     a user can't pick an impossible device."""
+    global _torch_load_error
     devices: list[str] = []
     try:
         import torch
@@ -41,8 +53,21 @@ def available_torch_devices() -> list[str]:
             devices.append("cuda")
         if getattr(torch.backends, "mps", None) and torch.backends.mps.is_available():
             devices.append("mps")
+        _torch_load_error = None
     except ImportError:
         pass
+    except Exception as exc:  # noqa: BLE001 -- any failure here must not stop the server
+        # Installed but unloadable: on Windows a CUDA DLL left behind by a
+        # half-reverted install raises OSError (WinError 127) from torch's own
+        # DLL loader (#723). Catching only ImportError let that kill startup
+        # and every /api/settings call. The server has to come up so it can
+        # say what is wrong. Logged loudly once per distinct cause, because
+        # separation cannot run on any device until torch is repaired, CPU
+        # included, and a quiet fallback would hide exactly that.
+        reason = f"{type(exc).__name__}: {exc}"
+        if reason != _torch_load_error:
+            logger.error("torch is installed but could not be loaded: %s", reason, exc_info=True)
+        _torch_load_error = reason
     devices.append("cpu")
     return devices
 

app/main.py

@@ -39,6 +39,7 @@
     available_torch_devices,
     configure_portable_environment,
     ensure_runtime_dirs,
+    torch_load_error,
 )
 from app.core.logging_setup import configure_logging
 from app.core.process import process_exists as _process_exists
@@ -393,6 +394,9 @@ def _settings_payload() -> dict[str, object]:
         "demucs_device": get_demucs_device_choice(),
         "demucs_device_resolved": get_demucs_device(),
         "demucs_devices_available": available_torch_devices(),
+        # Set when torch is installed but cannot load, which stops separation
+        # on every device (#730). Read after the probe above refreshed it.
+        "torch_error": torch_load_error(),
     }
 
 

app/pipeline/runner.py

@@ -21,7 +21,7 @@
 from app.core.redact import redact
 from app.core.registry import is_upload, set_proc
 from app.core.registry import persist as persist_registry
-from app.pipeline.analyze import analyze
+from app.pipeline.analyze import analyze, refine_key_from_stems
 from app.pipeline.beatgrid import compute_beat_grid
 from app.pipeline.collect import (
     cleanup_source,
@@ -222,6 +222,12 @@ def _run_common(job: Job, source: Path, job_dir: Path) -> None:
         cleanup_source(job_dir)
     job.stems = [{"name": name, "url": f"/api/jobs/{job.id}/stems/{name}.wav"} for name in found]
     _check_cancel(job)
+    # The key was first estimated from the mix, before separation. The stems
+    # are better evidence, the bass line above all (#726). Keeps the first
+    # estimate on any failure. Counted in "post", not as a stage of its own:
+    # the timings' order is a contract (tests/test_identify_regressions.py).
+    refine_key_from_stems(job, stems_dir)
+    _check_cancel(job)
     _set(job, stage="Mixing tracks...")
     original_path = make_original_track(job, job_dir, stems_dir)
     if original_path is not None: