Do not simplify in overloaded operators, type mappers

[inducer]

Closes #11.
Closes #149.

What finally tipped the balance is that the overload-y operators are hard to precisely type, which made typing various downstream user packages unnecessarily hard.

• Add a mapper that applies these same simplifications.
• Fix pymbolic CI
• Fix downstream CIs
• Downstream typing fixes
  • Pymbolic typing modepy#98
  • Pymbolic typing loopy#868
  • Port to new pymbolic pytential#242
  • pytato: WIP
• Finish typing at least the base mappers
• Figure out whether this can be compatible with present-day (unfixed) loopy enough to not break Firedrake
  • It cannot. See Prepare for non-simplifying pymbolic loopy#871.
• Reset ruff target version to 3.10
• Upgrade Python min version in dependencies, adjust CI versions
  • loopy
  • pytato
  • pytential
  • grudge
  • modepy
  • sumpy
• Updated required CI components

cc @kaushikcfd

[inducer]

@alexfikl This PR bumps the compat target to 3.10. The immediate reason was this inheritance, which is illegal in 3.9 and below. It can be hacked around, but it would have to be hacked around in a lot of places. 3.8 is EOL as of a few weeks ago, and I'm not sure I'm too enamored with supporting 3.9. Current numpy also has a 3.10 minimum. The net implication is that the whole stack will probably switch before too long. There's some ruff type annotation churn that results, but it's probably for the best to get that out of the way. How do you feel about that idea?

[alexfikl]

That sounds good to me. I've also jumped directly to 3.10 with most of my random code due to numpy bumping it. What's the conda default python these days?

[alexfikl]

Also, I'm not sure I've ever seen this before. What's IdentityMapper[[]] supposed to mean?

[inducer]

The corresponding type parameter on IdentityMapper is a ParamSpec for i.e. the extra arguments that get passed around by the mapper methods.

https://docs.python.org/3.13/library/typing.html#user-defined-generic-types says

The one exception to this is that a list of types can be used to substitute a ParamSpec

Based on the failures on 3.8 and 3.9, this must be a relatively recent thing.

[alexfikl]

Took a swift look over this and it looks good to me. Downstreams seem very upset with it though :(

[alexfikl]

Also, I'm not sure I've ever seen this before. What's IdentityMapper[[]] supposed to mean?

[inducer]

Downstreams seem very upset with it though :(

They sure are. I'm trying to figure out whether the damage can be controlled or whether we need to do this in a more piecemeal fashion.

[alexfikl]

pytential seems to work with this patch (didn't run all the testsuite though)

diff --git a/pymbolic/geometric_algebra/__init__.py b/pymbolic/geometric_algebra/__init__.py
index fbc7e30..4291b74 100644
--- a/pymbolic/geometric_algebra/__init__.py
+++ b/pymbolic/geometric_algebra/__init__.py
@@ -526,6 +526,7 @@ class MultiVector(Generic[CoeffT]):
"""

space: Space
+    mapper_method = "map_multivector"

# {{{ construction

diff --git a/pymbolic/mapper/coefficient.py b/pymbolic/mapper/coefficient.py
index 99516f0..b09b4f6 100644
--- a/pymbolic/mapper/coefficient.py
+++ b/pymbolic/mapper/coefficient.py
@@ -100,6 +100,9 @@ class CoefficientCollector(Mapper):
return {1: expr}

def map_constant(self, expr):
+        if expr == 0:
+            return {}
+
return {1: expr}

def map_algebraic_leaf(self, expr):

EDIT: Well, that was wrong, there are other errors too :(
EDIT: It seems to be missing some handling for MultiVector. Not sure why that worked before..
EDIT: Ugh, all the map_multivector need to be changed to map_multi_vector. Better to overwrite the default mapper_method for it?