Instead set their target-specific equivalents.
The POSITION_INDEPENDENT_CODE property isn't set because it already defaults to ON for shared libraries.
So that new warnings don't cause issues for library consumers.
Instead use CMake's CMAKE_COMPILE_WARNING_AS_ERROR option to treat them as errors in CI.
The tests all pass when using the bottom of the new version ranges on Windows.
The unicase dependency's lower bound is limited by rust-ini even after using the latest master branch versions of esplugin, libloadorder and loot-condition-interpreter.
The proc-macro2 and quote dependencies' lower bounds are limited by pyo3's dependency tree.
The nodejs wrapper's dependencies haven't been touched.
Long paths can be greater than 260 characters long, but each path component is still limited to 255 characters. The latter is also commonly the component length limit on Linux.
Support for long paths requires Windows 10 1607 or later and a Registry value to be set[1], so to make the cause of test failures clearer when that isn't the case, there's one test that checks that Registry value is set, and long paths are not used in the other tests if it isn't set, so only that one test should failed if the system isn't configured as expected.
GitHub Action's Windows runners do have the Registry value set[2].
This doesn't set the manifest for the Rust tests because it seems to be very difficult to set a manifest for only the tests, and it's not worth the effort when the only tests that fail are those that try to create a symlink, especially since those cases are also covered by the C++ tests.
[1]: https://learn.microsoft.com/en-us/windows/win32/fileio/maximum-file-path-limitation?tabs=registry#enable-long-paths-in-windows-10-version-1607-and-later
[2]: https://github.com/actions/runner-images/blob/releases/win22/20250921/images/windows/scripts/build/Configure-BaseImage.ps1#L73
The intent with the change to using shared_ptr in v0.27.0 was to make the caller responsible for their lifetimes, but I forgot to update the docs to reflect that, so this makes their lifetimes behave as documented.
Replace the boolean parameters for including user metadata and evaluating conditions with MergeMode and EvalMode enums.
The C++ and Python wrappers still use booleans because they're more constrained than enums in those languages, which can be given invalid values.
This makes it so that all non-public functions use camelCase and all public functions use PascalCase.
I regret choosing to use PascalCase for function names, but it's not worth breaking the whole public API to change that, and at least this approach has precedent in how Go decides if a function is public or private.
Oliver Hamlet