Hi Paul,
that expectation is the right one, and it is what I ended up fixing at the
source instead of in the test.
dart-sass 1.108.1-2, which I am about to upload, adds an arch:all package
"sass" that owns /usr/bin/sass. It is a three-line shell wrapper: it execs
/usr/bin/dart-sass where the native build exists (amd64, arm64, armhf,
riscv64), and runs the dart2js build of the same compiler, shipped in
node-sass, everywhere else. So the generic command is available again on
every architecture:
sass Architecture: all
Depends: dart-sass | node-sass
Why 1.108.1-1 did not have it: both packages can provide the compiler, but
dart-sass provides it ready-made, while the Node.js one has to be assembled by
hand here -- cli_pkg cannot run in a Debian build, so its bundle has to be
reproduced -- and it is more than twice as slow. Shipping the native command
and leaving the JavaScript package to bundlers was therefore the default.
What I had not realized is that dart-sass has an architecture gap, and that
gap turned "we provide the command" into "we provide the command on four
architectures". A bit more work and a bit slower is better than not having it
at all, so the solution is that wrapper: native where it exists, JavaScript
everywhere else.
Both builds come from the same Dart sources and accept the same options; the
only functional gap is --embedded, which the JavaScript build cannot do. I
verified the fallback on this machine (version output, compilation, --watch
through node-chokidar) and that upgrading from 1.108.1-1, where /usr/bin/sass
was a symlink in dart-sass, takes the path over cleanly.
So the test change is one line after all, just a different one:
-Depends: @, dart-sass
+Depends: @, sass
with no Architecture restriction, and openQA can make the same substitution
for its dependency on ruby-sass.
One caveat: "sass" is a new binary package name, so 1.108.1-2 goes through
NEW; until it is accepted the hint may still be needed. After that the
regression is gone.
Cheers,
Emfox