parser: clear arity error for deep prefix-binop chains

[danieljohnmorris]

Summary

qa-tester P1 rerun #6: +++a b c and other N-prefix-with-N-operands shapes used to exit silently with a bare ILO-P010 expected expression, got EOF at line 1 col 1. The agent had no signal about which chain was at fault or how to reshape it. The parser now surfaces an ILO-P003 pointing at the chain start with the actual glyph and the operand shortfall.

Manifesto framing: every silent-EOF on a real input is a retry the agent can't avoid, because there's no information to act on. A focused arity message lets the next call land first time, which is exactly the token-cost win the language is here for.

Repro

Before:

$ ilo 'main>n;a=1;b=2;c=3;+++a b c'
{"code":"ILO-P010","labels":[{"col":1,"end":0,"line":1,"message":"here","primary":true,"start":0}],"message":"expected expression, got EOF",...}

After:

$ ilo 'main>n;a=1;b=2;c=3;+++a b c'
{"code":"ILO-P003","labels":[{"col":20,...}],"message":"prefix chain `+++` needs 4 operands but found 3","suggestion":"a run of 3 prefix operators takes 4 operands. Add operands (e.g. `+++a b c d`), shorten the chain, or bind intermediate results first: `r=+a b;…r` keeps each step explicit."}

What's in the diff

• parser: detect deep prefix-binop chains with too few operands — adds prefix_chain_arity_diagnostic at the head of parse_prefix_binop, gated on scan_prefix_binary_end returning None so valid shapes (+a b, ++a b c, +++a b c d, +*a b c) stay on the existing path. The detector is conservative: only fires on K >= 2 leading prefix-binops, skips && / || (their dedicated hint still wins), and aborts on any unknown token in the operand area rather than risk a false positive. Counts unary wrappers (!~^$-) and atom postfix (.field, [i], (), !, with ...) so it doesn't miscount rich operands. Regression tests cover depths 2, 3, 4, a mixed-operator chain, and two positive shapes that must continue to parse cleanly.
• examples: prefix-chain-arity shows the K+1 operand rule — companion .ilo file so an agent that hits the new diagnostic has an in-context template for either fix (provide K+1 operands or bind intermediate results). Runs across every engine via the existing examples harness.

Test plan

• cargo test --release --features cranelift clean
• cargo clippy --release --features cranelift --all-targets -- -D warnings clean
• cargo fmt --check clean
• Original repro now emits ILO-P003 with chain glyph + shortfall
• Valid prefix forms (+a b, ++a b c, +++a b c d, +*a b c, &!x y, +p2.x p2.y) unchanged
• &&x y and ||x y still hit their dedicated hint, not the new one
• Examples harness exercises both deeparity and deepbind across all engines

Follow-ups

• The detector's K >= 2 gate keeps single-op +a shapes on the existing ILO-P010 path. If a future feedback round shows +a (missing second operand) is also a friction source, the same helper can drop the gate once operand-postfix accounting handles every shape parse_operand does.

[codecov]

Codecov Report

❌ Patch coverage is 65.24064% with 65 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	65.24%	65 Missing ⚠️

📢 Thoughts on this report? Let us know!