[v2.2.0] Cover extraction creates invalid files

[DigiBC]

Describe the bug

With a previous Liquidsoap development version from mid-March it was easily possible to extract cover art from metadata using the metadata.cover function. Besides MP3 and OGG metadata, FLAC metadata also worked (see #2952) and the extracted PNG and JPEG files were accurate.

In release version 2.2.0, the metadata.cover function only creates obviously invalid PNG and JPEG files that can't be displayed. Graphics software claim incorrect file headers.
Interestingly, the created files are somewhat larger than the original embedded PNG and JPEG cover art graphics.

To Reproduce

Unfortunately, I don't have a Liquidsoap developer version from that time anymore, but the following code example created accurate PNG files:

def extract_cover_art(m)
  file.write(append=false, data="#{metadata.cover(m)}", "/path/to/file.png")
end
source.on_metadata(extract_cover_art)

Expected behavior

Valid PNG and JPEG files that match the cover art graphics embedded as metadata.

Version details

• OS: Linux Mint 21.2 (Ubuntu 22.04 'Jammy') amd64
• Version: Liquidsoap 2.2.0

Install method

Install of official package 'liquidsoap_2.2.0-ubuntu-jammy-1_amd64.deb'
(also latest development packages)

[DigiBC]

In the meantime, I seem to have found the cause of the issue:
Shortly before the release of version 2.2.0, string recoding improvements were implemented in commit aa84675, referencing issues #3231 / #3238. As a result of this change, metadata is recoded to UTF-8 by default.
This makes a lot of sense, but unfortunately the current implementation also converts binary data such as cover art images embedded in the metadata.
This leads to invalid image files when extracting cover art from metadata.

That behavior doesn't occur in earlier Liquidsoap development versions. (The last artifact tested without issue was commit 31cac52.)

As an example, here's the header of an embedded JPEG file, in the original...

... and after the undesired UTF-8 recoding:

According to the help file strings_encoding.md there's a setting that could be used as a workaround for now:
settings.metadata.recode: set to false to prevent metadata from being converted to UTF-8.
However, this setting returns an error, probably because it hasn't been implemented yet. (It can't be found in the rest of the documentation either.)

[DigiBC]

After a few more tries, I finally seem to have found a workaround! 😄

Fortunately, the previously mentioned setting is available after all, but under a different name. So, the UTF-8 recoding of the metadata can be disabled as follows:
settings.request.metadata_decoders.recode := false

With this setting, cover art images will no longer be manipulated either, so the extracted PNG and JPEG files will be accurate and usable.

This workaround helps me for the time being, but that doesn't fix the issue that reasonable UTF-8 recoding shouldn't change binary image metadata.

[DigiBC]

I can confirm that the issue has been fixed! 👍
Thanks!