Cranelift JIT nil-sweep batch 3: arithmetic, comparison, numeric-unary

[danieljohnmorris]

Summary

Batch 3 of the Cranelift JIT-helper permissive-nil sweep. Routes the
arithmetic, comparison, and numeric unary/binary helpers through the
JIT_RUNTIME_ERROR TLS channel introduced in #254, with span_bits packed
at the cranelift call site so diagnostics render with a caret matching
tree and VM.

Batches 1 and 2 (#264 and #259) covered list/record/map helpers; this
batch is the next chunk down the residual-helpers audit list. Helpers
in scope (Group A, 21 functions):

• arithmetic: jit_add, jit_add_inplace, jit_sub, jit_mul, jit_div, jit_mod, jit_neg
• comparison: jit_gt, jit_lt, jit_ge, jit_le (EQ/NE keep 2-arg sig — NanVal-bit equality is total)
• numeric unary/binary: jit_abs, jit_min, jit_max, jit_flr, jit_cel, jit_rou, jit_clamp
• coercion / length: jit_len, jit_str, jit_num

Manifesto framing: arithmetic-slow-path divergence is the highest
token-cost-saved item in the residual sweep. Every silent nil from
n - "x" on a register the verifier can't prove numeric was a
bewildered agent retry. Now agents see the same cannot subtract non-numbers message they get on tree and VM, with the caret pointing
at the offending site.

Repro before / after

Before this change, exercising the type-error path of any Group A
helper (e.g. via the in-tree unit-test API that bypasses the verifier
to drive the helpers directly):

let r = jit_sub(str_val("hello"), num(1.0));  // old 2-arg sig
assert!(is_nil(r));                            // permissive-nil
// no error surfaced

After:

let _ = jit_take_runtime_error();
let r = jit_sub(str_val("hello"), num(1.0), 0);
assert!(is_nil(r));
let err = jit_take_runtime_error().expect("expected pending error");
assert!(matches!(err.0, VmError::Type(msg) if msg.contains("subtract")));

At the CLI level the verifier still rejects most surface programs
that statically mix types (ILO-T009 / ILO-T010 / ILO-T012), so the
slow-path helpers are usually unreachable from naive ilo. The error
plumbing matters for the dynamic-dispatch cases (FnRef returns, union
types via match results) and for any future codegen that emits the
slow path.

What's in the diff (per commit)

1. route arithmetic + comparison + numeric-unary JIT helpers through JIT_RUNTIME_ERROR

   • 21 helper signatures gain span_bits: u64 (or +1 for clamp's already-3-arg shape)
   • Each error path replaces bare TAG_NIL (or TAG_FALSE for ordered comparison) with jit_set_runtime_error_with_span(VmError::Type(...), span_bits) + TAG_NIL/TAG_FALSE
   • Error wording matches the VM dispatcher's vm_err! messages so cross-engine diagnostics are byte-identical
   • jit_len gains a Map arm to match tree/VM's "len requires string, list, or map" — was missing entirely
   • declare_helper arities updated in both src/vm/jit_cranelift.rs (in-process JIT) and src/vm/compile_cranelift.rs (AOT)
   • AOT dispatch sites pull pack_span_bits via the now-pub(super) helper
   • vm::tests::jit_helpers unit tests updated to assert the runtime-error cell is set and carries the right VmError variant

2. add cross-engine regression suite for batch-3 helpers

   • 29 tests asserting tree / VM / cranelift JIT happy-path parity for every helper touched in this batch
   • No-stale-error-leak guard pinning JitRuntimeErrorGuard's clear-on-entry contract

3. add example pinning batch-3 helper happy paths through examples_engines

   • examples/jit-nil-sweep-batch3.ilo with 24 entry points
   • Runs through tree / VM / cranelift via the examples_engines harness; doubles as in-context teaching example for agents

Test plan

• cargo build --release --features cranelift — green
• cargo clippy --release --features cranelift --all-targets — clean
• cargo fmt — clean
• cargo test --release --features cranelift --lib — 2932 passed, 0 failed (52 ignored)
• cargo test --release --features cranelift --test regression_jit_nil_sweep_batch3 — 29/29 passed
• cargo test --release --features cranelift --test examples_engines — passed
• Full cargo test --release --features cranelift — every suite green

Follow-ups

This is batch 3 of 7 in the Cranelift JIT-helper permissive-nil sweep.
Remaining batches queued (one PR each, separate worktrees):

• Batch 4 — Group B text + len/coerce: jit_fmt2, jit_trm, jit_upr, jit_lwr, jit_cap, jit_padl, jit_padr, jit_ord, jit_chr, jit_chars, jit_unq, jit_frq
• Batch 5 — Group C collections: jit_spl, jit_cat, jit_has, jit_range, jit_window, jit_zip, jit_chunks, jit_enumerate, jit_setunion, jit_setinter, jit_setdiff, jit_rev, jit_srt, jit_rsrt, jit_cumsum
• Batch 6 — Group D stats + linear algebra: jit_median, jit_quantile, jit_stdev, jit_variance, jit_fft, jit_ifft, jit_transpose, jit_matmul, jit_dot, jit_det, jit_inv, jit_solve
• Batch 7 — Group E I/O type-error paths: jit_rd, jit_rdl, jit_wr, jit_wrl, jit_jpar, jit_rdjl, jit_dtfmt, jit_dtparse

Per the audit, these helpers are correctly permissive-by-design and stay out of scope: jit_pow/jit_sqrt/jit_log/jit_exp/jit_sin/jit_cos/jit_tan/jit_log10/jit_log2/jit_atan2 (both VM and JIT return f64::NAN on non-numeric — already parity); jit_concat/jit_concat_inplace (only emitted for typed-string OP_ADD_SS by the type checker, defensive unreachable!() is correct); jit_listget OOB / jit_mget missing-key (legitimate-miss returns nil by O _ semantics).

[codecov]

Codecov Report

❌ Patch coverage is 92.98246% with 16 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	87.12%	13 Missing ⚠️
src/vm/compile_cranelift.rs	94.73%	3 Missing ⚠️

📢 Thoughts on this report? Let us know!