Unary minus on function arguments / let bindings folds incorrectly

[PearCoding]

Summary

Unary minus (`-`) applied to a function parameter or let-bound variable is miscompiled. In HikariRenderer's PExpr stdlib, `abs(-3.5)` returns `-3.5` instead of `3.5`.

Reproduction

Using the PExpr compiler embedded in HikariRenderer (`SVMCompiler`, `dev` branch of PExpr), with all optimisation passes enabled:

Expression	Expected	Actual
`abs(-3.5)`	`3.5`	`-3.5`
`abs(3.5)`	`3.5`	`3.5` ✓
`abs(-1.0)`	`1.0`	`-1.0`
`if (-3.5 < 0.0) { -(-3.5) } else { -3.5 }`	`3.5`	`3.5` ✓
`-(-3.5)`	`3.5`	`3.5` ✓
`let a = -3.5 in -a`	`3.5`	`0`
`let a = -3.5 in 0.0 - a`	`3.5`	`0`

The stdlib `abs` is defined as:
```
fn abs(a:num) = if a < 0 { -a } else { a };
```

The SSA dump for `abs(-3.5)` is a single line:
```
return -3.5:num
```
The `abs` call is fully folded away and the unary minus on the function parameter `-a` has been dropped.

Scope

• An explicit `if` with literal branches works — branch selection is fine.
• A top-level `-(-literal)` works — unary minus on a literal is fine.
• The bug shows up specifically when unary minus is applied to a function argument or a let-bound variable, i.e. when the operand is a reference to a binding being substituted/inlined.

Likely cause

Constant folding/inlining substitutes the argument's value into the function body but loses (or mis-handles) the surrounding unary minus. The `0 - a` path folding to `0` in the let case suggests `a` is being resolved to `0` when read from the binding — possibly an undefined/zero-initialised register slot for the let binding in one of the lowering passes.

Workaround

Use arithmetic negation instead of unary minus: `0.0 - x`. This also fails for let-bound variables (returns 0), so the workaround only applies to inlined literals.

Environment

• PExpr: `dev` branch as fetched by PearCoding/HikariRenderer (CPMAddPackage)
• Host compiler: MSVC 19.44, C++20
• OS: Windows 11 x64

Filed while building out CI-runnable unit tests for HikariRenderer's shader system.