Fix tree-walker escaping function from match-in-loop tail

[danieljohnmorris]

Summary

A ?match{} expression at the tail of an @ or wh loop body was silently early-returning the enclosing function with the match arm's value. Three personas independently reported this during assessment runs as "match arm silently exits the outer loop." VM and Cranelift were correct; only the tree-walker was wrong.

Engine-divergent silent miscompilation is the worst possible failure mode for an AI-targeted language: the program verifies, runs, returns the wrong number with no error signal, and the agent burns retries debugging logic that isn't broken. Same severity class as the slc/mset bug PR #150 fixed.

Repro

ilo 'f>n;cs=["1"];@c cs{rn=num c;?rn{^_:99;~v:42}};5' --run-tree f

Engine	Before	After
--run-tree	42 ❌	5 ✓
--run-vm	5 ✓	5 ✓
--run-cranelift	5 ✓	5 ✓

Root cause

Stmt::Match in eval_stmt had a special-case if is_last { return BodyResult::Return(v) } that fired whenever the match was the final statement of any body — function or loop. The function-body handler in eval_body already returns BodyResult::Value(last) for the function caller to convert into the function return, so the special-case at Stmt::Match was duplicating that logic incorrectly. Inside a loop body it made the match's value propagate as a function-level Return, bypassing whatever came after the loop.

What's in the diff

1. interpreter: tree-walker no longer escapes function from match-in-loop tail — one-line removal of the if is_last branch in Stmt::Match's Value(v) arm. is_last was the only consumer of that parameter and eval_body was the only caller of eval_stmt, so removed the plumbing entirely rather than leaving dead args that invite the special-case to come back.

2. tests + example: pin match-in-loop tail behaviour across engines — 6 cross-engine regression tests covering match-as-loop-tail, match-as-function-tail (sanity), side-effect arms, cnt/brk propagation, and the wh-loop analogue. examples/match-in-loop.ilo demonstrates both loop-tail and function-tail shapes with -- run: / -- out: directives so tests/examples_engines.rs exercises them across every engine.

Test plan

• cargo test --release --features cranelift — 3419 passed, 0 failed, 74 ignored (3413 baseline + 6 new)
• cargo fmt --all -- --check clean
• Repro returns identical output across tree, vm, cranelift
• Function-tail match (no surrounding loop) still returns the match's value correctly
• Code-reviewed by subagent; should-fix and nice-to-have findings addressed (full is_last plumbing removed, wh-loop test added, example expanded with function-tail sanity case, header comment trimmed)

Follow-ups

The doc entry about mset m k variable silently storing wrong value when a 0 was previously bound was a stale-binary artefact of PR #150 — same root cause, already fixed there. Updated ilo_assessment_feedback.md to reflect this in the Addressed section.

The custom ARM64 JIT removal is in flight on a separate branch (feature/remove-custom-jit) — independent of this PR, no conflict.

[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!