[ABA-23] test(vortex-array): repro for integer arith panics

[abnobdoss]

Summary

• Adds 5 inline repro tests in vortex-array/src/scalar_fn/fns/binary/numeric.rs (scalar_fn::fns::binary::numeric::tests) documenting the ABA-23 integer arithmetic panic bug.
• 4 tests are #[ignore] + #[should_panic(expected = "to_canonical failed")] — they lock the current wrong behaviour (panics on i32 add overflow, i64 div-by-zero, u32 sub underflow, i32 mul overflow).
• 1 test (issue_aba23_to_canonical_returns_err_not_panic_on_overflow) is always green: confirms to_canonical() (the VortexResult path) returns Err gracefully; only to_primitive() converts that Err to a panic via vortex_expect.

Linear: https://linear.app/abanoubdoss/issue/ABA-23

REPRO-ONLY PR — no fix included. The fix requires a cross-operator overflow-semantics decision.

Root cause (confirmed on develop)

ToCanonical::to_primitive() at canonical.rs:489-492 calls to_canonical().vortex_expect("to_canonical failed"). Arrow arithmetic kernels return Err on integer overflow/div-by-zero but evaluation is lazy — the error surfaces inside to_canonical where vortex_expect converts it to an unrecoverable panic. Callers cannot catch with VortexResult.

Open question — semantics decision needed

1. Return Err — to_primitive() propagates VortexResult; breaking API change. Matches Operator::Add doc.
2. Saturate — clamp to T::MAX/T::MIN. Silent but predictable.
3. Wrap — modular arithmetic. The numeric.rs:68 comment says "wrapping" but Arrow errors, not wraps.

Should this follow Arrow's checked_* convention universally, or be per-operator configurable?

Test plan

• cargo test -p vortex-array --lib scalar_fn::fns::binary::numeric::tests — 4 ignored, 1 green
• -- --include-ignored — all 5 pass (4 should_panic confirmed, 1 green)
• cargo +stable fmt -p vortex-array
• No regressions in existing mod test subtests

🤖 Generated with Claude Code