Add campaign layer, GPU model caching, file browser cleanup

- Campaigns: new table, CRUD API, React context + provider
- Sessions: scoped to campaigns, paths under campaigns/{id}/
- File browser: scoped per campaign, removed copy/paste/autoPlay
- Sidebar: campaign selector dropdown at top
- Transcribe: GPU model cached/released via job counter
- Jobs: status text updates dynamically in real-time
- Auto-redirect: blocked when summarize job is active

backend/app/jobs.py

@@ -1,6 +1,11 @@
 """
 Minimal background job runner.
+
+GPU jobs use a shared counter so the Whisper model stays cached between
+back-to-back transcriptions and is only released when no more GPU work
+is queued.
 """
+import threading
 import traceback
 from concurrent.futures import ThreadPoolExecutor
 
@@ -9,10 +14,23 @@ from .errors import PipelineError
 from .logging_config import get_logger
 
 log = get_logger(__name__)
-_executor = ThreadPoolExecutor(max_workers=2)  # transcription is GPU-bound anyway; no benefit to more workers
+_executor = ThreadPoolExecutor(max_workers=1)
+_gpu_jobs_queued = 0
+_gpu_lock = threading.Lock()
 
 
-def submit(job_id: str, fn, *args, **kwargs):
+def _unload_gpu():
+    """Release the cached Whisper model from GPU memory."""
+    from .pipeline.transcribe import unload_models
+    unload_models()
+
+
+def submit(job_id: str, fn, *args, requires_gpu=False, **kwargs):
+    if requires_gpu:
+        with _gpu_lock:
+            global _gpu_jobs_queued
+            _gpu_jobs_queued += 1
+
     def _run():
         log.info("Job %s starting", job_id)
         db.update_job(job_id, status="running", progress="Starting...")
@@ -37,5 +55,13 @@ def submit(job_id: str, fn, *args, **kwargs):
                 error=f"Unexpected error: {e}", error_stage="unknown",
                 error_detail=traceback.format_exc(),
             )
+        finally:
+            if requires_gpu:
+                with _gpu_lock:
+                    global _gpu_jobs_queued
+                    _gpu_jobs_queued -= 1
+                    if _gpu_jobs_queued <= 0:
+                        log.info("No more GPU jobs queued — unloading Whisper model")
+                        _unload_gpu()
 
     _executor.submit(_run)