Duplicate Script Run #49

A single button click can produce two concurrent Streamlit script runs, and when that happens resolve_backend("auto") can return different backends for the two runs: one cuml, one sklearn. The duplicate run is preempted before it writes session state, so results stay consistent, but it doubles GPU work and makes the log misleading.

Observation

Embed & Explore, A100 node, auto backend, 36 images:

09:17:10 Starting embedding generation      (x2)
09:17:12 Embeddings stored                  (x1, the other run was preempted)
09:17:31 Auto-resolved reduction backend to sklearn (CPU)
09:17:31 Auto-resolved reduction backend to cuML (GPU: NVIDIA A100-PCIE-40GB)
09:17:31 PROJECTION START                   (x2)
09:17:54 cuML UMAP subprocess completed; Projection complete   (this run wrote the results)
09:17:59 sklearn run: Dim reduction complete, never reaches _run_projection (preempted)

Cause

1. check_cuml_available() does a live import cuml on every call, with except ImportError: return False. Two threads importing cuML at the same instant is a CPython hazard: the second importer can hit a partially initialized package, raise ImportError, and be reported as "cuML not available". check_cuda_available() is cached; the cuML check is not.
2. The duplicate run itself: either a duplicated click event over a port-forwarded connection or a second browser tab. The log has no session id, so it cannot be told apart yet.