Remove custom ARM64 JIT backend

[danieljohnmorris]

Summary

Removes the custom ARM64 JIT (--run-jit) from ilo-lang. Keeps Cranelift JIT (--run-cranelift), register VM (--run-vm), and tree-walker (--run-tree).

The custom JIT was Apple-Silicon-only, hand-rolled AArch64 codegen, supported only ~12 numeric opcodes, had no CI coverage, and every dispatch fix (PRs #150 and #154 both) had to be mirrored into it separately. Through the manifesto lens it was pure maintenance burden with no agent-token benefit: it doesn't reduce tokens for any agent program (Cranelift handles the same opcodes and more, faster), it just multiplies our surface area for future fixes.

~1,148 net lines removed across 8 files, with no functional regression for any program ilo can run.

What's in the diff

1. vm: remove custom ARM64 JIT backend — deletes src/vm/jit_arm64.rs (627 lines) and the pub mod declaration; strips --run-jit from the CLI surface (Cli::args, Engine::Jit, dispatch arm, manual arg parser, help text); removes the arm64 bench code paths in run_bench; drops 13 run_jit: false defaults from unit-test struct literals. Gates jit_args/all_numeric behind #[cfg(feature = "llvm")] since LLVM is their only remaining consumer.

2. tests: drop ARM64 JIT-specific cases — removes the 19 run_jit_arm64_* tests in tests/eval_inline.rs and 2 run_jit_* tests in tests/cli_integration.rs. Their arithmetic-correctness coverage is already provided by the parallel Cranelift and VM test variants in the same files.

3. examples + test harness: remove engine-skip jit references — drops -- engine-skip: jit directives from 11 example files (now silent no-ops since the jit engine name is no longer iterated) and removes the unreachable engine.name == "jit" branch from tests/examples_engines.rs.

4. docs: relabel ilo JIT row to Cranelift in perf table — the row's string/record/file/api columns held real numbers the custom JIT cannot produce, so the figures were Cranelift's, mislabeled. Relabel rather than remove so the row continues to carry the JIT-tier perf comparison readers expect.

Breaking change

--run-jit is no longer a known flag. Acceptable for a 0.x language. Users wanting JIT compilation should use --run-cranelift (the default JIT engine), which is faster and supports the full opcode set.

Test plan

• cargo test --release --features cranelift — full suite green
• cargo fmt --all -- --check clean
• cargo clippy --all-targets -- -D warnings clean (default features)
• cargo clippy --all-targets --features cranelift -- -D warnings clean
• CLI spot-check: --run-tree, --run-vm, --run-cranelift all return correct output for f x:n>n;*x 2 with arg 5
• --run-jit is no longer accepted (returns arity error via trailing-var-arg, same behaviour as any unknown flag in that position)
• Code-reviewed by subagent; should-fix findings addressed (all_numeric cfg-gating, stale engine-skip cleanup)

Why now

Two recent PRs (#150 slc/mset, #154 friendly errors) each had to mirror dispatch changes into three engines: VM, Cranelift, and the custom ARM64 JIT. The custom JIT was the highest-friction surface and the lowest-leverage backend. Removing it makes every future ilo-lang fix smaller.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ All tests successful. No failed tests found.

📢 Thoughts on this report? Let us know!