test: cross-engine regression coverage for unq on number lists

[danieljohnmorris]

Summary

The 2026-05-11 edu-teacher persona report (lines 391 / 586 of ilo_assessment_feedback.md) said len unq xs hung with exit 137 / OOM on a L n of meaningful size. I went looking for the bug and could not reproduce it on main — at any size from [1,2,2,3] up to 100k all-unique elements, on any engine, in any path. The actual fix landed back in 5a30e00 (2026-03-06) which switched the list dedup path from raw-bits comparison to nanval_equal. That commit pre-dates the persona report by two months — they were almost certainly running a stale binary (their own entry #9 in the same batch documents ~/.cargo/bin/ilo being out of date and needing reinstall).

No code change here. The substantive fix is already in main. What was missing was lock-in coverage for unq on number lists specifically — tests/eval_inline.rs covered the text case but not the number-list case, so any future regression back to raw-bits comparison, RC mishandling, or quadratic blowup would slip through.

This PR adds that coverage across all three engines.

Repro

Before (claimed):

ilo 'f xs:L n>n;len unq xs' "[1,2,2,3]" → hang, exit 137

Now (verified on current main):

ilo 'f xs:L n>n;len unq xs' "[1,2,2,3]" → 3, exit 0, <50ms

Scaling check: 100k all-unique numbers → 100000, ~42ms. Linear in practice.

What's in the diff

Two commits, both additive:

1. test: cross-engine regression for unq on number lists — new tests/regression_unq_numbers.rs with 24 tests (8 cases × 3 engines): persona repro, order preservation, all-same, empty, all-unique, floats, negatives/zero, and a 1000-element stress test that catches quadratic-blowup OOM.

2. examples: unq on number lists — new examples/unq-numbers.ilo exercising the same shapes through the examples_engines harness for a second line of cross-engine coverage, and as an in-context example for future agents.

No production source touched.

Test plan

• cargo fmt clean
• cargo clippy --release --features cranelift --tests --all-targets clean
• cargo test --release --features cranelift --test regression_unq_numbers → 24/24 pass
• cargo test --release --features cranelift --test examples_engines → pass
• cargo test --release --features cranelift full suite green (2861 lib + all integration suites, 0 failures)

Follow-ups

The unq list path is O(n²) (inner out.iter().any(nanval_equal) scan) but scales linearly in practice up to 100k elements thanks to auto-vectorisation of the equality compare. A hash-keyed rework would be a future optimisation, not a fix — no live bug forcing it, manifesto win is theoretical. Deferred.

After merge, the entry will move out of 🛠 In Progress into ✅ Addressed citing 5a30e00 + this PR.

[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!