fix: sum and avg now work on VM and Cranelift, not just tree

[danieljohnmorris]

Summary

ilo file.ilo --run-vm and --run-cranelift reported Compile error: undefined function: sum / avg. The tree-walker had both builtins wired up directly; the VM compiler had dedicated arms for median, stdev, variance, quantile, and cumsum but none for sum or avg, so they fell through to the named-function lookup and errored.

Manifesto framing: every agent reaching for sum xs on the VM hits a confusing "undefined function" diagnostic that names a real builtin, then has to rewrite the call as fld + xs 0. Tokens wasted and trust burned. The fix restores symmetry with the rest of the stats family across all three engines.

While fixing this I noticed the existing stats family (median, stdev, variance, quantile) had a Cranelift correctness bug: the per-register F64 shadow used by the OP_*_NN fast paths was never refreshed after a helper call, so any arithmetic over a stats result silently produced 0 on --run-cranelift. Fixed in the same PR with cross-cutting coverage.

Repro

Before:

$ echo 'r>n;avg [1, 2, 3, 4]' > t.ilo
$ ilo t.ilo --run r
2.5
$ ilo t.ilo --run-vm r
Compile error: undefined function: avg
$ ilo t.ilo --run-cranelift r
Compile error: undefined function: avg

After: 2.5 on every engine. Empty list: sum [] returns 0, avg [] errors with avg: cannot average an empty list, matching tree-walker semantics on every backend.

What's in the diff

• ae3768c vm: new OP_SUM (165) and OP_AVG (166) opcodes with compile arms in build_call, dispatch handlers, vm_sum / vm_avg helpers, and jit_sum / jit_avg for the Cranelift handoff. Un-ignores vm_sum_basic, vm_sum_empty, vm_avg_basic now that the VM backs them.
• 393d7a8 cranelift: wire the new opcodes through the JIT and AOT compilers (helper FuncIds, declarations, symbol registrations, codegen blocks). In the same commit, refresh the F64 shadow after every stats helper call (median, quantile, stdev, variance, sum, avg) so subsequent OP_*_NN fast paths read the real numeric result instead of a stale zero. Both numeric-output classification passes (jit + AOT) include the new ops so reg_always_num flows through.
• 4f191c0 test: 16-test cross-engine regression file pinning the happy paths, the empty-list contracts (sum=0, avg=error), the wrong-type / non-numeric-element errors, and the F64-shadow hazard (including the pre-existing median variant). New examples/sum-avg.ilo picked up by examples_engines so every backend round-trips it.

Test plan

• cargo test --release --features cranelift --test regression_sum_avg_cross_engine — 16/16 pass
• cargo test --release --features cranelift --test examples_engines — pass
• cargo test --release --features cranelift --lib — pass (the only failures with CARGO_TARGET_DIR set are pre-existing AOT tests that look for libilo.a in the default target dir; clean target passes 7/7)
• Manual repro across all three engines (tree, VM, cranelift) for: sum/avg happy path, empty list, non-list arg, non-numeric element, arithmetic-on-results (the F64-shadow hazard)
• cargo fmt --check clean

Follow-ups

None for this fix. The OP_SUM/OP_AVG dispatch in the VM has the same inline shape as OP_MEDIAN; a future refactor could collapse the six stats codegen blocks behind a small helper, but that's pure cleanup with no behavioural change.

[codecov]

Codecov Report

❌ Patch coverage is 98.03922% with 3 lines in your changes missing coverage. Please review.
✅ All tests successful. No failed tests found.

Files with missing lines	Patch %	Lines
src/vm/mod.rs	97.10%	2 Missing ⚠️
src/vm/compile_cranelift.rs	97.14%	1 Missing ⚠️

📢 Thoughts on this report? Let us know!