If a previous launch set rotation=90 in QT_QPA_PLATFORM and the operator now picks 0° from the dropdown, the rotation= option must come back o

Pi 4 runs eglfs, which ignores the linuxfb ``:rotation=N`` plugin option (that silent no-op was the 2026.06.0 bug). eglfs reads QT_QPA_EGLFS_R

If a prior launch (or the Dockerfile) baked in a rotation, the helper must not double-append.

hx-on::after-request runs in global JS scope, so the bulk forms must NOT call Alpine component methods there (that throws ReferenceError and t

With shuffle_playlist on, the ~60s windowed-asset deadline cap must not reshuffle mid-cycle. Membership is unchanged, so the Scheduler should

No extension on the URL, but HEAD reports ``Content-Type: video/mp4`` → auto-download with the inferred extension.

HEAD reports ``Content-Type: text/html`` (a 404 page, a JSON error envelope's text/html landing page, ...) → stay as stream URL. The download

HEAD returns 4xx → stay as stream URL. Some origins respond 405 Method Not Allowed to HEAD; either way, downgrading to stream-mode keeps the c

Some HLS origins serve ``application/vnd.apple.mpegurl`` from URLs without a ``.m3u8`` extension. Reject those at the HEAD probe — downloading

Any network exception during the HEAD probe → stay as stream URL. The classifier is best-effort; we never block the create call on a flaky ori

The synchronous HEAD probe must run with the documented 5s timeout — operators are blocking on the POST /assets call. Any drift in the timeout

A URL whose path ends in a known single-file video container auto-downloads with the matching local extension. No HEAD call fires — extension

Non-http(s)/non-streaming schemes (file://, ftp://, ...) get the negative classify. The classifier deliberately refuses to download from anyth

HLS / DASH / SmoothStreaming manifests never auto-download — they describe a stream, not a single file. No HEAD call (the extension match shor

``rtsp://camera/feed.mp4`` is RTSP. Path extension does not promote it to an http(s) download.

RTSP / SRT / UDP / MMS are streaming-by-construction, even if the URL's path happens to end in ``.mp4``. The viewer plays them live via QtMult

cleanup() bails early if settings['assetdir'] doesn't exist.

Most ticks have no pending bounce — must not touch the browser; otherwise every asset_loop iteration would kill it.

The main-thread half of the handoff: when the flag is set, terminate the webview and clear current_browser_url so the next view_*() call respa

HEIC uploads route through ``normalize_image_asset`` on v1.2 / v2 only — image normalisation deliberately stays opt-in for the older endpoints

v1 / v1.1 deliberately do NOT route HEIC uploads through the image-normalize pipeline. Image normalisation is opt-in via v1.2 / v2 — promoting

POSTing a http(s) URL pointing at a single-file video container (.mp4 in this test) on v1.2 / v2 must rewrite the row to a local destination p

v1 / v1.1 deliberately do NOT auto-download remote video URLs — they preserve the literal-URL semantics older clients depend on, matching how

Stream URLs (HLS manifests, RTSP feeds, ...) must NOT be auto-downloaded — they describe a live stream rather than a single file. The serializ

Every API version that accepts a local video upload must dispatch ``normalize_video_asset`` after persisting the row. Regression test for GH

Test that v2 rejects video assets with non-zero duration.

Every API version that accepts ``mimetype='youtube_asset'`` must dispatch ``download_youtube_asset`` after persisting the row. v1.2 was previo

Switch the device to dd/mm/yyyy: the availability pickers must format AND parse in that shape, so saving a typed d/m/Y datetime lands the righ

A weekday-filtered, day-parted asset must render its weekday and time-window chips on the table row (12-hour clock default).

A windowed-but-currently-inactive asset must not poison the deadline with its long-past start_date — that would cause refresh_playlist() to fi

Django's default 4xx handlers render `<code>.html` from the project template path. This bypasses URL routing and goes straight at the handler

``Origin: https://anthias.local`` (port 443) posting to a server on ``Host: anthias.local:8000`` is a cross-origin request even though one sid

The static device-tree → subtype table is the source of truth for board detection. Any drift between this table and ``compute_envelope``'s exp

A host without ``/proc/device-tree/model`` (dev container, a non-DT bootloader, balena's restricted /proc) collapses cleanly to ``None`` inste

DRM card index is probe-order-dependent; HDMI-A-N mapping is stable.

Repeated identical results log at DEBUG; transitions re-log loudly. Guards against log spam since the helper runs on every play()/ set_asset(

If card0 hosts HDMI-A-2 and card1 hosts HDMI-A-1, HDMI-A-1 still wins. Guards against accidentally sorting on the full entry name (card0-...

Headless host (the test runner) has no /sys/class/drm cards or fb0 — function should return None cleanly so the server falls back to the confi

The kB count from /proc/meminfo is the source of truth for low-RAM gating. Tests pin the parsing against the kernel's actual formats (space- a

A truncated /proc/meminfo without a MemTotal line returns ``None`` so the caller treats RAM as unknown — same recovery path as a missing /proc

A non-integer in the MemTotal field returns ``None`` rather than crashing the host_agent unit.

A host where /proc/meminfo can't be opened (container, balena restricted /proc) returns ``None``.

Disabling auth flips ``auth_backend`` back to '' but keeps the User row intact so re-enabling later doesn't force a fresh password. Validating

``dispatch_normalize_video`` writes ``metadata.processing_started_at`` so the periodic reconciler can age the row. Regression guard for GH #28

A serializer class that doesn't carry ``_pending_normalize`` at all must not crash — defends against a future serializer that doesn't route th

A serializer that didn't flag normalisation (most uploads: JPEG/PNG/H.264) leaves the helper as a no-op — no spurious Celery dispatch.

Issue #2575's feedback-loop requirement: a failing CEC command must surface to the operator as a visible toast, not silently succeed or no-op.

The endpoint must fail fast (no 10 s subprocess) when neither /dev/cec0 nor /dev/vchiq exists. 503 telegraphs 'this server lacks the hardware

No /dev/cec0 or /dev/vchiq in the container by default — the experimental section must NOT render. Guards against accidentally surfacing CEC c

With a CEC device node present, both buttons render under an Experimental badge inside the System controls neighbourhood.

Local URIs hit the filesystem directly (cheap, no roundtrip to the server's view of the world).