DowngradeSublibraries has been failing on master since 2026-06-08 (runs on 06-08T11:48, 06-09T10:30, 06-09T14:40; example: https://github.com/SciML/OrdinaryDiffEq.jl/actions/runs/27251066249). The trigger timing matches #3728 adopting the strict allow_reresolve: false downgrade workflow from SciML/.github#71, the same change that exposed the root-project ADTypes conflict fixed in #3735 — but this workflow's failures have a different root cause and are not fixed by #3735.
The per-sublibrary jobs fail at the resolve step on minimal test-dependency conflicts. Example from lib/OrdinaryDiffEqCore: the downgrade pins DifferentiationInterface = 0.7.16, which excludes the NLSolversBase version range forced by NLsolve/DiffEqDevTools in the test target — unsatisfiable. Roughly 18 lib/* jobs fail in this pattern, each potentially with a different conflicting pair.
Fixing this needs a bound-raising pass over the sublibraries' [compat] entries for their test dependencies (raise each conflicting lower bound to the minimal mutually-resolvable version), analogous to what #3701 did for runtime deps. Until then, DowngradeSublibraries failures on PRs are pre-existing master failures, not PR regressions.
Investigation notes (logs, reproduction setup) are in https://github.com/ChrisRackauckas/InternalJunk/issues/41.
🤖 Generated with Claude Code
DowngradeSublibrarieshas been failing on master since 2026-06-08 (runs on 06-08T11:48, 06-09T10:30, 06-09T14:40; example: https://github.com/SciML/OrdinaryDiffEq.jl/actions/runs/27251066249). The trigger timing matches #3728 adopting the strictallow_reresolve: falsedowngrade workflow from SciML/.github#71, the same change that exposed the root-project ADTypes conflict fixed in #3735 — but this workflow's failures have a different root cause and are not fixed by #3735.The per-sublibrary jobs fail at the resolve step on minimal test-dependency conflicts. Example from
lib/OrdinaryDiffEqCore: the downgrade pinsDifferentiationInterface = 0.7.16, which excludes theNLSolversBaseversion range forced byNLsolve/DiffEqDevToolsin the test target — unsatisfiable. Roughly 18lib/*jobs fail in this pattern, each potentially with a different conflicting pair.Fixing this needs a bound-raising pass over the sublibraries'
[compat]entries for their test dependencies (raise each conflicting lower bound to the minimal mutually-resolvable version), analogous to what #3701 did for runtime deps. Until then, DowngradeSublibraries failures on PRs are pre-existing master failures, not PR regressions.Investigation notes (logs, reproduction setup) are in https://github.com/ChrisRackauckas/InternalJunk/issues/41.
🤖 Generated with Claude Code