hub.darcs.net :: darcs -> darcs-reviewed -> changes

This introduces the concepts (definitions) of positive/negative dependency and similarly for conflict. The invariants and proofs are now based on these notions. The proofs remain informal but should now be mathematically exact, i.e. formalizable in principle.

Otherwise we get errors when building the attoparsec dependency.

It doesn't make sense to output the complete patch info in a progress message and our progress machinery can't properly handle that anyway.

This allows building against current stackage nightly and restores building with ghc-9.8 which broke due to PVP violating updates of these packages.

This is the version ("resolver") that the debian haskell team currently uses to build haskell packages. To make darcs build against it (with no extra-deps), we also have to limit usage of tls>=2.0 to ghc>=9.8 (instead of ghc>=9.6) because lts-22.34 comes with ghc-9.6.6 but it still on tls-1.8.0.

From now on, whenever we make changes to our cabal file, we should be careful to check that - stack build still works (i.e. finds a valid build plan) and - we adapt our stack.yml to whatever the debian haskell team targets These two steps should ensure continued availability of darcs in debian.

The bug in the directory package that forces us to exclude these versions appears only on windows. This makes it possible to build darcs with stack against current stackage nightly, hopefully allowing darcs to get back into stackage and afterwards back into debian.

This is done by bundling (only) the single module as src/Darcs/Util/StrictIdentity.hs.

It now doesn't get quite as easily confused when you add tracing to dump internal data for debugging.

This is only so we can reconstruct failing QC test cases for named patches.

instead we'll have warnings in older GHCs

For now this is GHC 9.8

Shrinking patch names is problematic due to explicit dependencies and not needed if we generate only minimal, human readable (and always valid) names.

This is much cleaner than hacking the applied patch names into the existing repo models. Also led to a clean instance ArbitraryState (Named p) (which was buggy before) and may in the future allow to shrink Named patches by removing some of their explicit dependencies.