non-evaluator attribute requires manual propagation across entire call chain

Start by finding the original PR or issue that introduced the non-evaluator attribute, then survey its use outside mir-semantics, including the EXPERIMENT-no-llvm-kompile-updated branch. Trace how the Haskell backend sends call chains and requires clauses to the LLVM backend; done means identifying the root cause, choosing a solution direction, and implementing and verifying it.

Problem

The non-evaluator attribute marks a function so it is not evaluated by the LLVM backend. However, in practice, marking a single function as non-evaluator is not enough — every function in the call chain above it must also be manually marked, otherwise the Haskell backend will still send the caller to the LLVM backend, which then encounters a non-evaluator function it cannot handle.

Example:

syntax Int ::= A(Int) [function, non-evaluator]
syntax Int ::= B(Int) [function]

rule B(X) => A(X) +Int 1

• A is marked non-evaluator
• B is not marked non-evaluator
• When the Haskell backend simplifies B(42), since B is not non-evaluator, it sends B(42) to the LLVM backend
• The LLVM backend attempts to evaluate, unfolds the rule, and encounters A(42) — but A is non-evaluator and the LLVM backend cannot handle it correctly
• Workaround: manually add non-evaluator to B as well. But then all callers of B also need it, and so on — propagating up the entire call chain

The same issue occurs when a requires clause references a non-evaluator function.

This was discovered in the mir-semantics project (EXPERIMENT-no-llvm-kompile-updated branch), which has many functions unsuitable for LLVM backend execution.

Expected Behavior

non-evaluator should not require manual propagation. Possible directions:

• The compiler automatically analyzes the call graph and propagates the attribute
• The Haskell backend recognizes when a term contains non-evaluator functions and handles them itself
• The LLVM backend gracefully returns unevaluated terms when encountering non-evaluator functions

Suggested First Steps

Before choosing a solution, some investigation is needed:

1. Find the original PR/issue that introduced non-evaluator to understand the design intent
2. Survey non-evaluator usage in projects other than mir-semantics — do they encounter the same propagation problem?
3. Assess the semantics of non-evaluator in concrete execution (pure LLVM backend) scenarios (currently a potential concern, no observed issues yet)

Acceptance Criteria

• Research non-evaluator introduction history and usage across projects
• Identify root cause and solution direction
• Implement and verify