diff --git a/.github/workflows/test-runner.yml b/.github/workflows/test-runner.yml
index 297aae9b..695c5248 100644
--- a/.github/workflows/test-runner.yml
+++ b/.github/workflows/test-runner.yml
@@ -125,14 +125,18 @@ jobs:
         env:
           CODECOV_TOKEN: ${{ secrets.CODECOV_TOKEN }}
 
-      # Browser-test failure artifacts (Chrome screenshot + rendered
-      # HTML + console log) are written by tests/conftest.py's
-      # pytest_runtest_makereport hook into ./test-artifacts/. Same
+      # Browser-test failure artifacts (Playwright trace.zip +
+      # full-page screenshot per failed test) are written by
+      # pytest-playwright via the
+      # `--tracing retain-on-failure --screenshot only-on-failure
+      # --output test-artifacts` addopts in pyproject.toml. Same
       # bind-mount story as coverage.xml: anything inside the test
       # container under /usr/src/app shows up directly on the runner.
-      # `if: failure()` keeps green runs from uploading an empty bundle;
-      # `if-no-files-found: ignore` is a belt for the case where a
-      # failure is in setup/teardown and never reaches the hook.
+      # `if: failure()` keeps green runs from uploading an empty
+      # bundle; `if-no-files-found: ignore` is a belt for the case
+      # where the failure happens before any test reaches the
+      # tracing-stop call. `playwright show-trace trace.zip` replays
+      # the test interactively when downloading the bundle locally.
       - name: Upload integration test artifacts
         if: failure() && inputs.test-type == 'python'
         uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a # v7
diff --git a/src/anthias_server/app/views.py b/src/anthias_server/app/views.py
index 5ddd8b68..197df6c2 100644
--- a/src/anthias_server/app/views.py
+++ b/src/anthias_server/app/views.py
@@ -280,12 +280,14 @@ def assets_update(request: HttpRequest, asset_id: str) -> HttpResponse:
     # row marked as 'webpage', etc.). The edit modal renders mimetype
     # read-only for the same reason.
     if asset.mimetype == 'video':
-        # Video assets must keep duration=0 — the API enforces this on
-        # writes and the playlist scheduler reads the real length back
-        # from the file at playtime. The edit form disables the input
-        # for videos, but defend on the server too so a hand-crafted
-        # POST can't desync the DB from the API contract.
-        asset.duration = 0
+        # Video duration is owned by the probe_video_duration Celery
+        # task, not the edit form. Leave the persisted value alone so
+        # an edit doesn't clobber the probed length back to 0. The
+        # edit form already disables the duration input for videos
+        # (`:disabled="editAsset.mimetype === 'video'"`); this branch
+        # is the matching server-side guard against hand-crafted POSTs
+        # that try to write a duration anyway.
+        pass
     else:
         asset.duration = int(
             request.POST.get('duration') or asset.duration or 0
diff --git a/src/anthias_server/celery_tasks.py b/src/anthias_server/celery_tasks.py
index 58b1afe7..65b6c3e1 100755
--- a/src/anthias_server/celery_tasks.py
+++ b/src/anthias_server/celery_tasks.py
@@ -8,7 +8,7 @@ from typing import Any
 
 import django
 import sh
-from celery import Celery
+from celery import Celery, Task
 from tenacity import Retrying, stop_after_attempt, wait_fixed
 
 django.setup()
@@ -205,7 +205,47 @@ def cleanup() -> None:
             logging.warning('cleanup: could not remove %s: %s', entry.path, e)
 
 
+class _ProbeVideoTask(Task):  # type: ignore[type-arg]
+    """Custom Task subclass so ``on_failure`` can clear
+    ``is_processing`` when retries are exhausted.
+
+    Without this, a row whose probe permanently fails (e.g. ffprobe
+    missing on a stripped-down image, or 3 consecutive ffprobe
+    timeouts) stays at "Processing" forever and the operator has no
+    way to interact with it. Celery calls ``on_failure`` after
+    ``max_retries`` is exhausted *or* on any non-autoretry exception
+    that escapes the task body.
+    """
+
+    def on_failure(
+        self,
+        exc: BaseException,
+        task_id: str,
+        args: tuple[Any, ...],
+        kwargs: dict[str, Any],
+        einfo: Any,
+    ) -> None:
+        asset_id = args[0] if args else kwargs.get('asset_id')
+        if not asset_id:
+            return
+        try:
+            Asset.objects.filter(asset_id=asset_id).update(is_processing=False)
+            # Same WS nudge the success path sends so the row drops
+            # the "Processing" pill without waiting for the 5s table
+            # poll. Operator then has the row in its terminal state
+            # and can edit/delete it.
+            from anthias_server.app.consumers import notify_asset_update
+
+            notify_asset_update(asset_id)
+        except Exception:
+            logging.exception(
+                'probe_video_duration on_failure cleanup failed for %s',
+                asset_id,
+            )
+
+
 @celery.task(
+    base=_ProbeVideoTask,
     time_limit=120,
     autoretry_for=(sh.TimeoutException, sh.ErrorReturnCode, OSError),
     retry_backoff=10,
@@ -233,6 +273,9 @@ def probe_video_duration(asset_id: str) -> None:
         is_processing=False so the operator can adjust manually.
       - Any other exception is swallowed (logged) so a programmer
         error in the helper doesn't hammer the worker via retries.
+      - Retries exhausted (autoretry_for raise after max_retries) →
+        _ProbeVideoTask.on_failure clears is_processing so the row
+        doesn't stay stuck at "Processing" indefinitely.
     """
     try:
         asset = Asset.objects.get(asset_id=asset_id)