I used it on a project with many thousands of files and hundreds of (mostly dynamic) build rules. This was a literate program, and most build steps required the extraction of code from the source docs --- including build rules themselves --- and I still maintained sub-second updates.
I think Joe would have loved it.
In my experiments though, Tup failed hard on point 2, lacking a good language to factorize the dependency graph declaration. (and to be clear, I think Make's is pretty terrible by itself, before GMSL or Guile extensions come into the picture)
Edit: I see tup supports Lua extensions, which may cancel my complaint above.
What language and/or system were you writing in? I discovered literate programming in 2003 and was excited by it (still am). But it never took off and the tooling was poor (Leo, the Literate Programming Editor, is the one I remember - and shudder thinking about)
https://news.ycombinator.com/item?id=12398098
I'm still on that "detour," but still a believer.
Edit: some dead links there. This "bootstrap" directory contains the custom tangler and the custom rule processor. This 10K, plus the tiny configs in the root, are the only code in the project not in documents.
https://bitbucket.org/gavinpc/willshake/src/master/bootstrap...
I did look at SCons years ago, which is a Python-based build-system, but as I recall it ended up being the worse of both worlds by attempting a declarative syntax that was marred by its imperative warts.
and your points sound like you're just describing a dsl in lisp/scheme.
At it's core, it's really a mix of three things:
1. A language to define a dependency graph ("target: dependencies") and how the nodes are connected ("rules")
2. a functional scripting language to be able to factorize the declaration of the graph
3. and finally a graph traversal worker to dispatch and perform those rules that are required (including a sophisticated "job server" to maintain the concurrent job limit across child make processes with no developer intervention!)
I haven't encountered a good replacement tool that properly tackles all 3 of these feature sets. (Granted, I haven't done an in-depth analysis of the field...)