perf(vm): swap Box for Rc in Function::Memoized inner function 🧠

[timfennis]

Summary

• Function::Memoized held the inner function as Box<Self>, so every clone of the variant (one per memoized callsite via resolve_callee) deep-cloned the boxed function — a heap alloc + recursive Function::clone on every call.
• Switch to Rc<Self>: outer clones become refcount bumps, and the inner is only materialized on cache miss via Rc::unwrap_or_clone.
• Identity semantics are unchanged — they were already pointer-based on the cache Rc.

Benchmarks

Targeted micro-bench: 10M memoized calls, 99% cache hits, minimal body work.

Workload	Baseline	Patched	Speedup
pure fn f(x) { x + 1 } (compiled, no captures)	1.656s ± 0.021s	1.566s ± 0.019s	1.06×
let f = pure fn(x) { x + bias } (closure)	1.713s ± 0.025s	1.593s ± 0.038s	1.08×

10 runs each, 3 warmup. The closure case sees a slightly bigger win because the old Box clone walked a Vec<Rc<UpvalueCell>>; the new path just bumps a single Rc.

On a real workload (AoC 2022 day 16, a memoized DFS), the per-call savings drown in the body's own work and the result is within noise — but the change is never a loss, and it's faster wherever memoization dispatch is the bottleneck.

Test plan

• cargo test — all 300+ functional tests pass
• cargo fmt, cargo clippy -p ndc_vm — no new warnings
• Output of both microbenchmarks matches between baseline and patched builds

🤖 Generated with Claude Code