[AdaWorldAPI]

Two independent commits.

1. 7257896 — re-exports through ndarray::simd

Plain pub use, no wrappers, no behaviour change:

• hpc::vml: vsexp, vsln, vssqrt, vsabs, vsadd, vsmul, vsdiv, vssin, vscos, vspow, vstanh, vserf, vsfloor, vsceil, vsround, vsneg, vstrunc
• hpc::activations: sigmoid_f32, softmax_f32, log_softmax_f32, softmax_axis_f32, log_softmax_axis_f32
• hpc::amx_matmul: matmul_f32_batched, MatmulError, gated #[cfg(target_arch = "x86_64")] like the module

matmul_f32 itself is not re-exported: under runtime-dispatch the name simd::matmul_f32 already belongs to the simd_runtime trampoline, which has a different signature.

Tests: tests/simd_facade_reexports.rs (4) checks each simd:: path gives bit-identical output to its hpc:: home. Disable run: dropping vserf from the re-export fails the build with E0425.

2. df04b87 — hpc::checked_i64 (the Tract-requested int64 wishlist)

Signatures:

pub enum I64Op { Add, Sub, Mul, Div }
pub enum I64ArithError {
    Overflow { op: I64Op, position: usize, lhs: i64, rhs: i64 },
    DivisionByZero { position: usize, lhs: i64 },
    LengthMismatch { lhs: usize, rhs: usize, out: usize },
    RangeZeroDelta,
    RangeExceedsCapacity { len: u64, capacity: usize },
}
pub fn checked_i64(op: I64Op, lhs: i64, rhs: i64) -> Result<i64, I64ArithError>;
pub fn checked_binary_i64(op: I64Op, lhs: &[i64], rhs: &[i64], out: &mut [i64]) -> Result<(), I64ArithError>;
pub fn range_len_i64(start: i64, limit: i64, delta: i64) -> Result<u64, I64ArithError>;
pub fn range_i64(start: i64, limit: i64, delta: i64, out: &mut [i64]) -> Result<usize, I64ArithError>;

Policy, shared by every operation through checked_i64:

• A result outside i64 is refused with the operation, output position and operands. Never wrapped, never saturated.
• Div truncates toward zero (ORT integer division). Zero divisor → DivisionByZero; i64::MIN / -1 → Overflow.
• Operands are flat slices of the output's length or length 1 (scalar broadcast). Any other length → LengthMismatch. N-d broadcasting stays in Tract's graph.
• On any refusal out is unchanged: every element is checked before out is written.
• Range follows ONNX: max(ceil((limit - start) / delta), 0) elements, computed exactly in i128. A range running the wrong way is empty. delta == 0 → RangeZeroDelta. A range longer than out → RangeExceedsCapacity (the bound is the caller's buffer). The full i64::MIN..i64::MAX span reports len = u64::MAX.

Negative fixtures (all in negative_fixtures_refuse_with_their_information and its neighbours):

input	refusal
Add(i64::MAX, 1)	Overflow { op: Add, position: 0, lhs: i64::MAX, rhs: 1 }
Sub(i64::MIN, 1)	Overflow { op: Sub, .. }
Mul(3_037_000_500, 3_037_000_500)	Overflow { op: Mul, .. }
Mul(i64::MIN, -1)	Overflow { op: Mul, .. }
Div(i64::MIN, -1)	Overflow { op: Div, .. }
Div(5, 0)	DivisionByZero { position: 0, lhs: 5 }
Add([1, 2, i64::MAX, 4], [1])	Overflow { position: 2, .. }, out unchanged
Div([6, 6, 6, 6], [3, 2, 1, 0])	DivisionByZero { position: 3, lhs: 6 }, out unchanged
lengths 2 and 4 into 4	LengthMismatch { lhs: 2, rhs: 4, out: 4 }
Range(0, 10, 0)	RangeZeroDelta
Range(0, 5, 1) into 4 slots	RangeExceedsCapacity { len: 5, capacity: 4 }, out unchanged

Tests (7 unit + 3 doctests), all pass on x86_64 and i686:

• every pair from a 13-value edge set (including i64::MIN, i64::MAX, ±√2⁶³ and ±2³²) × 4 ops against exact i128 arithmetic: each result either matches, or is refused exactly when the exact value does not fit;
• truncation toward zero, distinguished from floor division (-7 / 2 = -3, not -4);
• Range against a reference loop over start, limit ∈ [-9, 9] and 7 deltas, plus both ends of i64.

Disable runs, each red then restored:

disable	failing test
D1 Add wraps	3 tests
D2 zero-divisor check removed	3 tests
D3 check-before-write pass removed	slices_broadcast_… (out written on refusal)
D4 position not reported	slices_broadcast_…
D5 length check weakened	unsupported_broadcasting_refuses
D6 delta == 0 accepted	range_refuses_zero_delta_and_overlong_output
D7 range truncated to capacity	range_refuses_zero_delta_and_overlong_output

No shape graph, no ONNX interpreter, no new generic types. Not re-exported through ndarray::simd: this is plain scalar arithmetic, not SIMD.

Checks

• cargo clippy --features std --all-targets -D warnings: clean.
• aarch64 cargo check --tests: clean (commit 1).
• --features std,runtime-dispatch: clean, no name collision.
• FACET STRIDE guard on the diff: clean.

Not merged automatically.

🤖 Generated with Claude Code

https://claude.ai/code/session_01XAqKMPh2Seh2WLccCksAdY

[coderabbitai]

Important

• 🔍 Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

Run configuration

• Source mode: Auto
• Repository: Unavailable
• Branch: Unavailable
• File: No configuration file from this repository applied
• Commit: Unavailable
• Layers: Organization UI
• Review profile: CHILL
• Plan: Advanced
• Run ID: 26d42511-71f1-4044-9b3e-439db69d46b6

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-30T11:28:05.704513Z	df04b87	Draft marked ready

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.