``_format_subprocess_stderr`` must produce operator-readable text: bytes decoded as UTF-8 (with replacement for malformed bytes), no ``b'...'`

A .tmp younger than 1h is mid-upload and must survive.

In-flight yt-dlp sidecars (<1h) must survive the sweep.

End-to-end run: 3 assets, middle one fails. After the queue drains, the header reads "Migration finished with errors" (not the all-clear copy)

If ASSET_X is active and ASSET_X[end_date] == (now + 3) and ASSET_TOMORROW will be active tomorrow then deadline should be ASSET_TOMORROW[star

No host_agent (balena fleets) or an empty publish falls back to reading the device tree in-container — this is what upgrades the ``anthias-roc

A host_agent-published value wins; the device tree is not consulted when Redis already answers.

Redis empty + unknown device tree → ``None`` (caller keeps the raw DEVICE_TYPE and the conservative empty arm64 codec set).

The Celery task persists the CEC power state to redis as a str. diagnostics.get_display_power() returns ``str | bool``; redis-py raises ``Dat

Every failure path in this module is degraded through gracefully (backoff + cached verdict), so none of it belongs at ERROR level or above — t

The stock ``Fast 1080p30`` preset caps output at 1080p, so it doubles as the low-RAM resolution fix — there's no separate Dimensions / Resolut

An empty supported set (unrecognised arm64 board) has no transcode target, so there are no HandBrake steps to offer — matching ``_ffmpeg_reenc

An H.264 board's walkthrough leans entirely on the stock ``Fast 1080p30`` preset (which already outputs H.264 MP4), so it must NOT tell the op

A Pi 5 (HEVC only — no H.264 in its set) can't use the preset's default H.264, so the walkthrough adds a Video-tab step switching the encoder

An unreachable api.github.com is a routine condition on a signage device (offline installs, locked-down networks) — it must not log at ERROR,

The no-response branch must log the exception detail, not a bare 'no data' — Sentry previously showed 'ConnectionError fetching latest release

Issue #2856 (Copilot review #2): if wlr-randr can't apply the transform (cage not ready, no outputs, every output failed), we must NOT latch t

Two ``reload`` messages back-to-back with the same rotation must not flap the pending-bounce flag or set the skip_event twice on an already-pe

Most reload traffic (asset edits, etc.) must NOT blank the screen — the rotation-change path is keyed on a genuine delta.

The dark-mode flag is only read at webview launch, so a live toggle has to respawn AnthiasViewer. Like the rotation path, _handle_reload runs

linuxfb only reads :rotation=N at QPA init, so the live-rotation path has to bounce AnthiasViewer. _handle_reload runs on the subscriber threa

The reload pub/sub message must re-push wlr-randr when the operator changes rotation in Settings — that's the whole point of issue #2856's UI-

On linuxfb (Pi) the rotation change DOES need MediaPlayerProxy reset, because VLC bakes the transform filter into the instance at construction

``reload`` must still reload settings — that path is exercised by the settings.patch() endpoint and predates the skip behaviour.

Sanity: home loads, the homeApp x-data root mounts, no 5xx.

A six-row, mixed-mimetype schedule must render the asset table with every row's drag handle and action cluster reachable. The per-row tests be

Common reverse-proxy shape: the upstream ``Host`` carries an explicit (often default) port — e.g. ``Host: anthias.local:443`` — while the brow