fix: split -N v into subtract at fresh-expression positions

[danieljohnmorris]

Summary

Manifesto framing: this is a token-cost fix. Six personas in the assessment log (L705, L1252, L1465, L1503, L1545, L1650, L1809) hit the same papercut three or more times each writing numerical formulas. The natural unary-negation idiom -0 v (meaning 0 - v) silently produced wrong results because Logos's -?[0-9]+... regex greedily consumed the leading -, so -0 v lexed as Number(-0) + stray Ref(v). The canonical workaround - 0 v (with space) works but is easy to forget when transcribing maths, and the failure mode is silent. Every retry costs tokens for everyone downstream.

Repro

Before:

$ ilo "ab x:n>n;-0 x" ab 7
{"code":"ILO-P001","message":"unexpected token Ref(\"x\")...","severity":"error"}

After:

$ ilo "ab x:n>n;-0 x" ab 7
-7
$ ilo "ab x:n>n;-0 x" ab -7
7

What's in the diff

Three commits:

1. lexer: split glued negative literal back into Minus + Number - new post-lex pass in src/lexer/mod.rs that rewrites Number(-N) into Minus, Number(N) only when the preceding token is one that introduces a fresh expression position: start of input, ;, \n, =, {, or (. The parser's existing parse_minus resolves the rest. Call-arg negatives (at xs -1, +a -3, into -3 0 10, <r -0.05, [1 -2 3]) keep their Number(-N) token. LBracket is deliberately excluded so [-2 1 3] stays a 3-element list. 10 new lexer unit tests pin every split context and every keep-literal case.

2. test: cross-engine regression coverage for neg-literal papercut - new tests/regression_neg_literal_papercut.rs exercises five fixed-shape programs (one per split context) plus three keep-literal pins, each across tree, VM, and Cranelift.

3. examples: neg-literal-papercut.ilo for the now-correct behaviour - five small functions (ab, absp, diff, zarm, neg) with -- run: / -- out: annotations. Doubles as a cross-engine regression via the examples_engines harness and as in-context learning material for future agents.

Test plan

• cargo test --release --features cranelift clean across tree/VM/Cranelift
• cargo fmt --check clean
• cargo clippy --release --features cranelift --all-targets -- -D warnings clean
• 10 new lexer unit tests pin all split + keep contexts
• 3 new cross-engine regression tests in regression_neg_literal_papercut.rs
• examples_engines exercises the new neg-literal-papercut.ilo across all engines
• Existing regression_negative_literal_after_op still green (call-arg literals preserved)

Follow-ups

None. The split rule is narrow by design - any expansion (e.g. also splitting after Comma for function calls that use comma separators, or after prefix-binop tokens) would need its own design pass and cross-engine coverage. If a persona hits a different fresh-expression context that I missed, adding it to the matches! arm in src/lexer/mod.rs:415 is a one-line change with a corresponding unit test.

[codecov]

Codecov Report

❌ Patch coverage is 98.31933% with 2 lines in your changes missing coverage. Please review.
✅ All tests successful. No failed tests found.

Files with missing lines	Patch %	Lines
src/lexer/mod.rs	98.31%	2 Missing ⚠️

📢 Thoughts on this report? Let us know!