LibZ | LibZ , the alternative to ILMerge

In the error message, we can see

If you see errors about a specific library missing, you can use dnf itself to find out the name. For example, on Fedora 35:

You don't need to use pip to install packages on repl.it -- and in fact, you shouldn't! Using Nix derivations not only works better (as you're using their OS distro the way it's designed), but also keeps their storage costs low, by allowing packages to be used from a read-only, hash-addressed, shared store.

Binaries built for other distributions might assume that there will be libraries in /lib, /usr/lib, or the like, but that's not how NixOS works: Libraries will be in a path like /nix/store/--/lib, and those paths get embedded into the executables that use those libraries.

The easiest thing to do here is to create a new bash repl, but to add a Python interpreter to it. (I suggest this instead of using a Python repl because the way they have their Python REPLs set up adds a bunch of extra tools that need to be reconfigured; a bash repl keeps it simple).

• Create a new bash repl.
• Click on the three-dots menu.
• Select "Show Hidden Files".
• Open the file named replit.nix
• Edit the file by adding a Python interpreter with pandas, as follows:

For future reference, the error was indeed in the cmake process detecting wrong info regarding libpng. Adding set(CMAKE_FIND_FRAMEWORK LAST) in the CMakeLists.txt fixed the issue! As reported at OSX + homebrew + CMake + libpng version mismatch issue by Kikaxa:

The problem is that classical cmake-style Find*.cmake search for the headers and libraries separately - and the results MAY and WILL mismatch in some cases. MacOS exaggerates the problem by having special case of frameworks, being searched BEFORE other locations by default.

OK, so the issue here was the way dlopen link the libraries.
For my case, updating dlopen from
dlopen(so_lib_name, RTLD_NOW)
to
dlopen(so_lib_name, RTLD_NOW|RTLD_GLOBAL)
worked!

After lots of frustration it became apparent that the problem was once again in front of the computer. The following option in my git config was the culprit:

You will need to install the libcurl-devel package as it contains the headers files you are missing.

The libcurl-devel package includes header files and libraries necessary for developing programs which use the libcurl library. It contains the API documentation of the library, too

Git requires libcurl for HTTP and HTTPS support. Because a Git without HTTPS support is not very useful, most distro packages of Git depend on the distro package of libcurl.

What is likely happening in your case is that when you install Git, its dependency, libcurl, is also being installed from the distro package. The custom libcurl is either installed in the normal location without a suitable package and is therefore being overwritten, or it's installed somewhere later in the library search path and is being ignored because the distro version is being preferred.

You should run ldd $(which curl) both before and after the installation of Git and find out where the custom and distro versions of libcurl are located. If the image is installing libcurl into somewhere under /usr/lib, then it's broken: that location is reserved for the package manager, and it would need to build and install a normal system package with its custom version to avoid being overwritten. Otherwise, you may be able to modify /etc/ld.so.conf to adjust the search path of the library, or use the LD_LIBRARY_PATH environment variable to do so on a per-program basis.

The solution I found was to cross-compile with the toolchain from raspberrypi.org and set --sysroot= to a Raspbian image. Their releases more or less follow Debian's and always support all existing Raspberry Pies.
Both CXX and LD point to their g++, meaning it's GCC who determines linking flags. If necessary I can use -Wl,..
Other than -Wabi-tag -D_GLIBCXX_USE_CXX11_ABI=0, CFLAGS is pretty standard.

Raspbian changed their download URL format between Buster 10.3 (2020-02-13) and 10.4 (2020-05-28) and started providing full and lite versions. I used the latter.
Since these images typically have almost no free space, I first extend the image file, partition (p2) and underlying filesystem, so I can install the -dev packages that satisfy the dependencies of my project by chrooting into the sysroot via quemu-arm.