add matvec builtin for matrix-vector product

[danieljohnmorris]

Summary

Adds matvec xm ys > L n, a native matrix-vector product. Skips the wrap-as-column-matrix + flatten ceremony every linear-regression-style persona was paying:

flat (matmul xm (map (y:n>L n;[y]) ys))   ->   matvec xm ys

Three lines, ~10 tokens per use-site, gone. The linear-regression persona that surfaced this would have its flatten (matmul (transpose xm) (map (y:n>L n;[y]) ys)) recipe collapse to matvec (transpose xm) ys.

Closes pending.md #5an (P1+S).

Repro before / after

Before (manual recipe):

out = flat (matmul [[1,2,3],[4,5,6]] (map (y:n>L n;[y]) [7,8,9]))

After:

out = matvec [[1,2,3],[4,5,6]] [7,8,9]

Both produce [50, 122]; the regression suite asserts they stay equal.

What's in the diff

Single logical commit:

• Builtin::Matvec variant + name/from-name/ALL (appended to preserve every existing on-wire tag) + builtin-name round-trip test arrays.
• Verifier: ("matvec", &["L (L n)", "L n"], "L n") table row + a per-arg ILO-T013 arm.
• Interpreter: #[inline(never)] fn matvec_run next to matrix_from_value / vec_from_value. Errors return ILO-R009 on empty matrix, zero-cols, ragged rows, and dim mismatch. The call_function arm is a single tail-call into the helper, matching the lstsq feature: add lstsq builtin for ordinary least squares #515 / feature: crypto primitives - sha256, hmac-sha256, base64, hex, ct-eq #494 / feat(builtins): URL + base64url encoding cluster (urlenc/urldec/b64u/b64u-dec) #506 pattern that keeps the dispatch frame off cargo-nextest's tighter stack budget.
• VM: (Builtin::Matvec, 2) => true in is_tree_bridge_eligible, so VM and Cranelift JIT/AOT inherit semantics with no new opcodes.
• Docs: SPEC.md linalg table row + linalg prose section, ai.txt (auto-regenerated by build.rs), skills/ilo/SKILL.md 4-char+ list, new Linear algebra section in skills/ilo/ilo-builtins-math.md, CHANGELOG Unreleased Added entry.
• Tests: 9 cross-engine cases in tests/regression_matvec.rs (identity, 2x2, 2x3, 1x1, negatives+zeros, parity vs flat matmul ceremony, dim mismatch, empty, ragged). VM + Cranelift JIT; error tests are VM-only (matches the existing regression_linalg_basic.rs convention since the JIT path bubbles tree-bridge errors as nil).
• examples/matvec.ilo worked example with -- run: / -- out: markers across every engine via tests/examples_engines.rs.

Test plan

• cargo test --release --features cranelift --test regression_matvec — 9/9 pass
• cargo test --release --features cranelift --test examples_engines — pass (matvec example runs through every engine)
• cargo test --release --features cranelift full suite — pass (after the standard libilo.a symlink workaround for the AOT tests, pending #5av in flight as Claude/playwright ilo exploration r opnx #63)
• cargo fmt --check clean
• cargo clippy --release --features cranelift --all-targets -- -D warnings clean
• Verifier rejects matvec [1,2,3] [4,5,6] (first arg not L (L n)) with ILO-T013
• Verifier rejects matvec [[1,2],[3,4]] [[5,6]] (second arg not L n) with ILO-T013

Follow-ups

• When split call_function into 14 inline-never family dispatchers #504 lands and decomposes call_function, the #[inline(never)] band-aid on matvec_run can come off alongside every other pre-emptive helper.

[codecov]

Codecov Report

❌ Patch coverage is 90.80460% with 8 lines in your changes missing coverage. Please review.
✅ All tests successful. No failed tests found.

Files with missing lines	Patch %	Lines
src/interpreter/mod.rs	91.66%	4 Missing ⚠️
src/verify.rs	88.23%	4 Missing ⚠️

📢 Thoughts on this report? Let us know!