commit 78de6c9a3dd6c9909b16be503d4eb6777ef06f37
Merge: fead187 e924fe1
Author: Juan Manuel Méndez Rey <juan.mendezr@proton.me>
Date: Fri Oct 2 16:06:49 2026 +0000
Merge branch 'vcs-dart-team' into 'debian/latest'
Point Vcs-* at the Debian Dart Team
See merge request dart-team/dart-sdk!1
commit e924fe1597038ff6146d194acd0132b9f77dba8d
Author: Juan Manuel Méndez Rey <juan.mendezr@proton.me>
Date: Thu Oct 1 19:19:17 2026 +0200
Point Vcs-* at the Debian Dart Team
The package was moved into dart-team on salsa, so debian/control still
pointed at what is now only a fork. No changelog entry: this rides along
with whatever goes up next.
commit fead187b8a2da258732fda20291d686495aef4b1
Author: Juan Manuel Méndez Rey <juan.mendezr@proton.me>
Date: Thu Oct 1 15:16:20 2026 +0200
Shorten the -6 changelog entry
commit 3078f37fb45ceccf158dad1d33326784745f48f2
Author: Juan Manuel Méndez Rey <juan.mendezr@proton.me>
Date: Thu Oct 1 15:14:41 2026 +0200
Prepare 3.13.4+dfsg1-6, source-only, for reproducibility
All four architectures now have binaries in unstable, so the buildds can
take this one and it goes up source-only.
commit e048ebb7a0c5405fe80df9d23644f880d1a54c87
Author: Juan Manuel Méndez Rey <juan.mendezr@proton.me>
Date: Tue Sep 29 14:08:51 2026 +0200
Build the app-jit snapshots as kernel snapshots, for reproducibility
britney refuses to migrate this package to testing with "New but not
reproduced on amd64", and the age requirement is already satisfied, so the
reproducibility check is a hard block rather than an advisory one. Three
app-jit snapshots differ between builds because they are produced by a
training run that records whatever the JIT chose to optimise;
analysis_server.dart.snapshot came out at four different sizes across four
builds.
dart_snapshot_kind=kernel is the value upstream already selects for cross
builds. With it, two builds using the rebuilder's own build path produce a
byte-identical .deb and dbgsym, and autopkgtest passes. That also settles
the question of whether the app-jit noise was hiding a second source of
nondeterminism: it was not.
The cost is about 0.24 s on every "dart run", measured as interleaved
medians on an idle machine, and nothing else: the analyzer, the IDE path,
dart compile exe and dart compile js never open these snapshots. The
package also loses 36 MB.
This is temporary. Upstream is removing app-jit snapshots altogether in
https://github.com/dart-lang/sdk/issues/63978, which fixes this without
costing any startup time, and the setting should be dropped when that
lands.
No changelog entry yet: the sponsor has an NMU in flight, so the next
version number and whether it has to acknowledge that NMU are not settled.
commit fafab8c9bfe990bd79cd21c6211dea86ca0cc0dd
Author: Juan Manuel Méndez Rey <juan.mendezr@proton.me>
Date: Tue Sep 29 08:44:23 2026 +0200
README.source: point the reproducibility note at upstream #63978
Upstream closed the report as a duplicate of the work to remove app-jit
snapshots from the SDK entirely, which they hope to finish for Dart 4.
Their position is that making the training runs deterministic is not worth
doing for packaging alone, and they are right that the problem disappears
once nothing is trained. Record that, so the kernel-snapshot workaround is
not reached for by someone who has not read why it costs 0.6 s per
dart run.
Among the 4 debian patches available in version 3.13.4+dfsg1-6 of the package, we noticed the following issues:
![[bug history graph]](https://qa.debian.org/data/bts/graphs/d/dart-sdk.png){kind=link}