parser: reject reserved stmt keywords as fn decl starts

[danieljohnmorris]

Summary

Both gis-analyst and routing-tsp personas hit the same parser trap on v0.11.1: wh >cond{...} mid-body gets re-parsed as a new function declaration named wh returning cond, surfacing as a misleading ILO-T008 return type mismatch: expected n, got _ plus a cascade of "undefined variable" errors in function wh. Same root cause as the v0.10 wh >len xs 0{...} friction entry — it was never len-specific, just wh > ... in any position.

Manifesto framing: the workaround (cnd=>v 0;wh cnd{...;cnd=>v 0}) adds two extra statements per loop, exactly the kind of recurring token tax the language exists to avoid. The fix dissolves the friction at the parser layer; the corrected wh >v 0{...} form is 1 token cheaper per loop than any workaround.

Repro

Before:

$ ilo run 'foo s:n>n;v=+s 0;wh >v 0{v=- v 1};+v 0' foo 5
error[ILO-T008]: return type mismatch: expected n, got _
error[ILO-T004]: undefined variable 'v' (in function 'wh')
error[ILO-T004]: undefined variable 'v' (in function 'wh')

After:

$ ilo run 'foo s:n>n;v=+s 0;wh >v 0{v=- v 1};+v 0' foo 5
0

What's in the diff

Commit 1 — parser: reject reserved stmt keywords as fn decl starts. is_fn_decl_start previously accepted any Ident Greater ... sequence as a zero-param fn header. After the first ; of an outer body, parse_body_with's top-level boundary heuristic fires on wh >v 0{...} and treats wh as a fresh fn decl. Short-circuit is_fn_decl_start to false when the leading ident is one of the four reserved statement-keyword identifiers (wh/ret/brk/cnt) that parse_stmt intercepts as control-flow forms. Mirrors the existing ILO-P011 reserved-keyword precedent in parse_decl.

Commit 2 — test: cross-engine coverage for wh >cond parser trap. New tests/regression_wh_gt_trap.rs covers four shapes across tree, vm, and cranelift:

• wh >v 0{...} after a let binding (the gis-analyst repro)
• wh >cond{...} followed by a sibling fn decl (boundary heuristic must still find the real boundary)
• wh >=v 1{...} (GreaterEq prefix, same family)
• a zero-param fn sanity check (legitimate name>return;body still parses)

examples/wh-gt-condition.ilo demonstrates the now-correct shape and rides through tests/examples_engines.rs as an extra cross-engine regression layer.

Test plan

• cargo test --release --features cranelift — 4253 passed, 0 failed
• All 12 new regression_wh_gt_trap tests pass on tree/vm/cranelift
• examples_engines picks up wh-gt-condition.ilo and runs it across all engines
• cargo fmt --check clean
• cargo clippy --release --features cranelift --all-targets -- -D warnings clean
• Headline repro foo s:n>n;v=+s 0;wh >v 0{v=- v 1};+v 0 returns 0 for foo 5 under all three engines

[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!