fix: prefix-binop operand expands known-arity calls

[danieljohnmorris]

Summary

Six-rerun P1 standing: wh >len q 0{...} failed on first attempt across every persona (gis-analyst, nlp-engineer, others), forcing a bind-first or paren retry every time.

Root cause: parse_prefix_binop parsed both operands via parse_operand (atom-only), so >len q 0 parsed as BinOp(>, Ref(len), Ref(q)) and orphaned 0 — fatal in a wh header where the surrounding context immediately expects { and emits ILO-P003.

Fix mirrors the prefix-?? swap from #310: new parse_prefix_binop_operand helper dispatches to parse_call_arg when the ident at the cursor is a known-arity function AND the next token can start another operand. Falls back to parse_operand otherwise, so bare locals that shadow user fn names (&e f{...} where f is a local) keep resolving via Ref instead of expanding into a zero-arg call. The shadow-safety guard is what kept the FizzBuzz regression (vm_and_does_not_clobber_left_operand) green.

Manifesto win: the natural shape parses on first try. No more "wh prefix-cond" retries chewing through tokens.

Repro

Before:

$ echo 'main>n;q=[1,2,3];wh >len q 0{q=tl q};len q' | ilo run /dev/stdin
{"code":"ILO-P003","message":"expected LBrace, got Number(0.0)",...}

After:

$ echo 'main>n;q=[1,2,3];wh >len q 0{q=tl q};len q' | ilo run /dev/stdin
0

What's in the diff

• parser: prefix-binop operand expands known-arity calls (commit 1) — parse_prefix_binop now delegates each operand to parse_prefix_binop_operand, which opts into parse_call_arg only when the next ident is known-arity AND followed by an operand-startable token. Suppression set (records, field access, zero-arg paren, postfix-bang) mirrors the existing one in parse_call_arg.
• tests: cross-engine regression for prefix-binop with call operands (commit 2) — tests/regression_prefix_binop_call.rs covers the originating wh >len q 0 shape plus guard form, let-RHS, prefix-ternary, equality, both-operands-as-calls, parenthesised right operand. Pins the bare-local fallback (wh >v 0 keeps working). Tree + VM + Cranelift.
• examples: wh-prefix-call shapes for the engine harness (commit 3) — examples/wh-prefix-call.ilo runs through examples_engines.rs so the working shapes are in-context examples for future agents and the example harness catches any regression independently of the parser-specific test.

Notable behaviour change

+f g where f is a 1-arity user fn now parses as BinOp(+, Call(f, [g]), <next>) instead of BinOp(+, Ref(f), Ref(g)). The bare-ref shape was never meaningful (you can't add two function references), so this is a clean upgrade — consistent with the same trade ?? made in #310. If a real program breaks it's an easy bind-first rewrite.

Test plan

• cargo test --release --features cranelift — full suite green
• New regression_prefix_binop_call tree/VM/Cranelift all pass
• examples_engines harness picks up the new example
• Previously-failing vm_and_does_not_clobber_left_operand (FizzBuzz fixture) still green after the shadow-safe guard
• regression_wh_gt_trap (existing wh >v 0 cases) still green
• regression_prefix_nil_coalesce (the precedent fix from fix: prefix ?? accepts call expression as value side #310) still green

Follow-ups

None for this PR. The analogous prefix-ternary (?>len q 0 a b) is already covered by the new tests and works because prefix-ternary's condition flows through parse_prefix_binop. parse_prefix_ternary's then/else operands still use parse_operand — that's a separate trade-off worth opening only if a persona hits it.

[codecov]

Codecov Report

❌ Patch coverage is 78.12500% with 7 lines in your changes missing coverage. Please review.
✅ All tests successful. No failed tests found.

Files with missing lines	Patch %	Lines
src/parser/mod.rs	78.12%	7 Missing ⚠️

📢 Thoughts on this report? Let us know!