negative indices on slc / take / drop (parity with at)

[danieljohnmorris]

Summary

PR #183 added Python-style negative indices to at xs i last week. This PR extends the same rule to the slice-shaped builtins so "negative index = count from the end" is uniform across every position-taking primitive:

• slc xs s e — negative s and e both resolved against len, clamped to [0, len]. slc xs -1 (len xs) is the last element as a 1-element list; slc xs 0 -1 drops the last element.
• take n xs — negative n means "all but the last |n|", i.e. Python's xs[:n]. take -1 [1,2,3] returns [1,2].
• drop n xs — negative n means "keep only the last |n|", i.e. Python's xs[n:]. drop -1 [1,2,3] returns [3].

Closes the quant-trader fencepost (at eq np ILO-R009 because @i 1..np loops produce np-1 results), and the older slc xs -np 1 np ergonomics gap. One rule everywhere; no last xs builtin (the token-economics math doesn't justify a new primitive when at xs -1 is 3 tokens and already documented in examples/at-indexing.ilo).

Repro

Before:

xs=[10,20,30,40,50]
slc xs -1 5    # ILO-R009 / silent wrap on usize cast
take -1 xs     # ILO-R009 "take: count must be a non-negative integer"
drop -1 xs     # ILO-R009 "drop: count must be a non-negative integer"

After:

slc xs -1 5    # [50]
take -1 xs     # [10, 20, 30, 40]
drop -1 xs     # [50]

Bounds clamp; never wrap, never error on -99.

What's in the diff

Four causally-linked commits:

1. builtins: shared helpers — resolve_slice_bound, resolve_take_count, resolve_drop_count defined once with 8 unit tests covering positive / -1 / -len / beyond-len / empty boundaries. Lets the three engines dispatch through identical logic, same pattern PR at: Python-style negative indexing #183 established with char_at_signed.
2. interpreter: tree-walker — slc, take, drop routed through the helpers. slc picks up the same fract != 0 integer check take/drop and at already had (it previously cast floats silently).
3. vm: OP_SLC/OP_TAKE/OP_DROP + jit_slc/jit_take/jit_drop — same routing on VM bytecode and Cranelift JIT helpers. Refreshes the stale regression_take_drop.rs asserts that locked in the old "non-negative" errors plus the cranelift TAG_NIL workaround.
4. test + example — 28 cross-engine tests in regression_neg_index_slice.rs exercising the matrix (3 engines × -1/-len/beyond-len/0/len × list/text) plus the quant-trader fencepost as a named test, plus examples/negative-indices.ilo so the engine harness exercises six canonical patterns and agents have a worked reference in-tree.

One deliberate tightening worth flagging: slc indices now require fract == 0. Previously the tree-walker and VM both silently cast non-integer floats like 1.5 to usize. This brings slc to parity with take, drop, and at, and the diagnostic is precise ("slc: start index must be an integer"). Vanishingly unlikely to break a real program but called out so anyone hitting it sees the deliberate boundary.

Test plan

• cargo test --release --features cranelift — full suite green (the one transient AOT-test flake is pre-existing — the vm::compile_cranelift::tests::aot_* tests share /tmp/ilo_test_aot_* filenames; passes in isolation).
• 28 new cross-engine tests in regression_neg_index_slice.rs cover slc / take / drop on lists and text, including -1, -len, beyond--len, unicode codepoints, empty-list edge case, and the quant-trader fencepost.
• Updated regression_take_drop.rs locks the new behaviour across all three engines (previously asserted error on tree/VM and nil on cranelift — that drift is gone).
• examples/negative-indices.ilo covers six idioms exercised by examples_engines.rs on every engine.
• cargo clippy --release --features cranelift --all-targets -- -D warnings clean.
• cargo fmt --check clean.

Follow-ups

• The slc xs -np 1 np parse friction in the assessment doc (line 753 of ilo_assessment_feedback.md) is mostly a -np lexer/parser issue separate from the slc semantics fixed here. Now that the semantics are right, that entry's workaround pattern goes away: a user can write slc xs 0 -1 directly without binding s=- np 1 first.

[codecov]

Codecov Report

❌ Patch coverage is 86.76471% with 18 lines in your changes missing coverage. Please review.
✅ All tests successful. No failed tests found.

Files with missing lines	Patch %	Lines
src/vm/mod.rs	75.60%	10 Missing ⚠️
src/interpreter/mod.rs	73.33%	8 Missing ⚠️

📢 Thoughts on this report? Let us know!