fix: clarify kebab-case is atomic in undefined-variable hint

[danieljohnmorris]

Summary

Persona reports (lines 503, 622, 958 in ilo_assessment_feedback.md) kept claiming best-d in str best-d parses as str(best) - d when best and d are both bound, eroding the kebab-case identifier win. The reports came in steadily enough that they read like a real precedence bug.

It isn't. The lexer regex is [a-z][a-z0-9]*(-[a-z0-9]+)* with logos longest-match at priority 1, so best-d is always one Ident token. The parser never sees Minus between the halves and cannot produce BinOp::Subtract. The cost the persona was actually paying was the error message:

ILO-T004: undefined variable 'best-d'
hint: did you mean 'best'?

"Did you mean 'best'" reads exactly like "we split your identifier and only the left half resolved." So a model that hits this error mid-task concludes "kebab-case is broken in str-call position," paws around for a workaround, and logs the bug again next session.

This PR is diagnostic-layer only - no grammar surgery. When an undefined kebab-case identifier has every dash-separated half resolving as a scoped variable, function, or builtin, the hint now says:

'best-d' is a single identifier (kebab-case); for subtraction write '- best d'

For 3+ segments the explicit subtraction form is dropped (there's no single binary-subtract spelling for a-b-c). When only one half resolves the standard closest-match suggestion stays. Token cost of the kill-the-confusion-at-the-source path is much lower than the recurring "log it, find the workaround, document the workaround" loop.

Repro

$ ilo "f>t;best=10;d=3;str best-d" f

Before:

ILO-T004 undefined variable 'best-d'
hint: did you mean 'best'?

After:

ILO-T004 undefined variable 'best-d'
hint: 'best-d' is a single identifier (kebab-case); for subtraction write '- best d'

When the kebab ident IS bound, output is unchanged (99, not 7) - the parser was always doing the right thing here, only the error path was misleading.

What's in the diff

• verify: clarify kebab-case in undefined-variable hint - new kebab_subtract_hint helper next to closest_match, wired into the ILO-T004 site for Expr::Ref in infer_expr. Falls back to the existing closest-match suggestion when the kebab theory doesn't apply.
• tests: cross-engine regression for kebab-case lexer and diagnostic - new tests/regression_kebab_precedence.rs. Five positive tests pin the lexer guarantee in str-call arg, list element, multi-segment, digit-segment, and explicit-subtraction forms across tree/VM/Cranelift. Three diagnostic tests cover the new hint, the 3-segment no-subtract-form variant, and the closest-match fallback.
• examples: kebab-vs-subtract, three forms side by side - new examples/kebab-vs-subtract.ilo exercised by examples_engines.rs. Same scope binds best, d, and best-d; [str best-d, str best, str d] prints [99, 10, 3] to make the atomicity obvious to any future agent that lands in the file.

Test plan

• cargo test --release --features cranelift passes (full suite, 0 failures)
• cargo fmt --all -- --check clean
• cargo clippy --release --features cranelift --all-targets -- -D warnings clean
• Manual repro of all three diagnostic paths (kebab-both-halves, 3-segment, closest-match fallback)
• Manual confirmation that str best-d with best-d bound still returns the kebab value, not subtraction
• examples/kebab-vs-subtract.ilo runs green via the engine harness on tree, VM, and Cranelift

Follow-ups

• Move persona entries at lines 503, 622, 958 of ilo_assessment_feedback.md to the canonical ✅ Addressed section after merge, with a one-line note clarifying the diagnostic-layer root cause.

[codecov]

Codecov Report

❌ Patch coverage is 96.77419% with 1 line in your changes missing coverage. Please review.
✅ All tests successful. No failed tests found.

Files with missing lines	Patch %	Lines
src/verify.rs	96.77%	1 Missing ⚠️

📢 Thoughts on this report? Let us know!