feat: ✨ live Dask dashboard for single-GPU runs

On a GPU run with no existing client, start a 1-worker/1-thread
in-process LocalCluster: GPU evals stay serial (no VRAM contention)
but a Dask dashboard is exposed and its URL logged, so progress is
visible. Falls back to the threaded scheduler if distributed/bokeh are
absent; a user-provided cluster is used as-is.

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

Author: lguerard

docs/guide/performance.md

@@ -17,6 +17,22 @@ merge step are sized to the host automatically:
 The RAM figure is read live via `psutil`; without it, a conservative default is
 used instead of guessing high.
 
+## Live progress dashboard (GPU runs)
+
+A single-GPU run still gets a **Dask dashboard**: patchworks spins up a tiny
+1-worker / 1-thread in-process cluster, which keeps GPU evaluations serial (no
+VRAM contention) while exposing the dashboard so you can watch tiles stream
+through. The URL is logged at the start of staging:
+
+```text
+INFO:patchworks._core:Dask dashboard for this run: http://127.0.0.1:8787/status
+```
+
+This needs `distributed` (and `bokeh` for the UI) installed; if they are
+missing, patchworks logs a warning and falls back to the threaded scheduler
+(no dashboard, same result). A cluster you start yourself
+(`make_local_cluster`) is used as-is instead.
+
 ## Overriding the worker count
 
 ```python

src/patchworks/_core.py

@@ -332,7 +332,31 @@ def active_fn(block, block_info=None):
     import dask as _dask
 
     _tile_nbytes = int(np.prod(labeled.chunksize)) * labeled.dtype.itemsize
-    if _active is None:
+    _temp_cluster = None
+    _temp_client = None
+    if _active is None and use_gpu:
+        # Single-GPU runs still get a live Dask dashboard: a 1-worker /
+        # 1-thread in-process cluster keeps GPU evals serial (no VRAM
+        # contention) while exposing the dashboard for progress.
+        try:
+            from dask.distributed import Client, LocalCluster
+
+            _temp_cluster = LocalCluster(
+                n_workers=1, threads_per_worker=1, processes=False
+            )
+            _temp_client = Client(_temp_cluster)
+            logger.info(
+                "Dask dashboard for this run: %s",
+                _temp_client.dashboard_link,
+            )
+        except Exception as exc:  # no distributed/bokeh → threaded fallback
+            logger.warning(
+                "Could not start a dashboard cluster (%s); "
+                "falling back to the threaded scheduler.",
+                exc,
+            )
+
+    if _distributed_client() is None:
         _workers = (
             max_workers
             if max_workers is not None
@@ -363,6 +387,9 @@ def active_fn(block, block_info=None):
     logger.info("Staging tiles to %s …", stage_path)
     with _sched_ctx:
         _stage_to_zarr(labeled, stage_path, "staged", progress)
+    if _temp_client is not None:
+        _temp_client.close()
+        _temp_cluster.close()
     labeled = da.from_zarr(stage_path, component="staged")
 
     # NB: no post-staging skip-count pass here — counting skipped tiles by