Codegen is not run-to-run deterministic: string-constant numbering, function-name registration order, and method-callee selection permute

[proggeramlug]

Compiling the same source twice with the same perry binary produces different LLVM IR. Not a miscompile — the programs behave identically — but it defeats any byte-level IR A/B, which is the primary evidence a rooting-campaign slice offers (#7615).

Repro

cargo build --release -p perry -p perry-runtime-static -p perry-stdlib-static
for run in 1 2; do
  mkdir -p /tmp/det/$run && cd /tmp/det/$run
  env PERRY_RS4GC=0 PERRY_GC_MOVING_LOOP_POLLS=1 PERRY_INLINE_SHADOW_SLOT=0 \
      PERRY_NO_AUTO_OPTIMIZE=1 \
    perry compile test-files/test_gap_object_create_method_this.ts -o out --trace llvm
done
diff /tmp/det/1/.perry-trace/llvm/*.ll /tmp/det/2/.perry-trace/llvm/*.ll

Two independent shapes, both reproducible:

1. String-constant numbering and function-name registration order.

< @.str.3 = private unnamed_addr constant [9 x i8] c"describe\00"
< @.str.4 = private unnamed_addr constant [9 x i8] c"getState\00"
< @.str.5 = private unnamed_addr constant [5 x i8] c"wrap\00"
---
> @.str.3 = private unnamed_addr constant [5 x i8] c"wrap\00"
> @.str.4 = private unnamed_addr constant [9 x i8] c"describe\00"
> @.str.5 = private unnamed_addr constant [9 x i8] c"getState\00"

with the matching permutation of js_register_function_name calls in __perry_init_strings_*_chunk0.

2. Method-callee selection permutes. In test_gap_class_expr_dynamic_parent_ctor, three consecutive dispatches swap which perry_method_* symbol each call site names:

<   %r2492 = call double @perry_method_..._DerA__promise(double %r2956)
>   %r2492 = call double @perry_method_..___anon_class_17__promise(double %r2956)
<   %r2493 = call double @perry_method_..___anon_class_17__promise(double %r2955)
>   %r2493 = call double @perry_method_..._DerB__promise(double %r2955)

Both smell like iteration over a HashMap/HashSet whose order is randomised per process.

Why it matters beyond tidiness

• It masquerades as a regression. Comparing IR between two compilers is how the Layer 1 slices demonstrate "root plumbing only". In refactor(codegen): migrate arrays_finds.rs + array_methods.rs onto the Layer 1 rooting API (#7615) #7620 three of nine differing functions were this, not the change — distinguishable only because the instruction multiset was identical and a same-binary double compile reproduced the same permutation. A slice that used diff would have reported three phantom differences; one that used a multiset comparison and did not check determinism would have quietly dismissed a real ordering change.
• .perry-cache keys on codegen env vars (build: the object cache doesn't key on codegen env vars — PERRY_WRITE_BARRIERS=0 silently does nothing with a warm cache #6394), not on output. A nondeterministic emitter means two cache entries for the same inputs can differ.
• Reproducible builds and any future "did this refactor change codegen?" gate both need this closed.

Filed from #7620; not fixed there.