[shinaoka]

Summary

Refs #355.

• Demote only the S2 timing ratio to a report: shared-runner machine variability makes it unsuitable as a hard gate. Keep the 1.5 reference and all timing, ratio, and allocation columns visible.
• Preserve the strict zero-allocation gate for both S2 paths; leave other rows and their gates unchanged.
• Add a deterministic regression test for ratio and allocation decisions.

Local validation

• OVERHEAD_GATE=1 taskset -c 0-7 cargo test -j 4 -p strided-fused --benches: passed; S2 reported with zero allocations.
• Temporarily changed only the S2 reference to 0.01: OVERHEAD_GATE=1 taskset -c 0-7 cargo test -j 4 -p strided-fused --bench overhead_smoke succeeded with S2 ratio 1.118 and report (ref 0.01; alloc=0); restored 1.5 afterward.
• cargo test -j 4 --workspace, parallel-feature tests for basic/kernel/fused, both release overhead smoke gates, cargo doc --workspace --no-deps, formatting and repository-rules review passed.
• cargo clippy -j 4 -p strided-fused --all-targets --no-deps passed with existing warnings. Workspace-wide clippy on rustc 1.99 fails on the existing clippy::uninit_vec at strided-view/src/view.rs:1440; workspace -D warnings also stops at existing clippy::suspicious_arithmetic_impl in strided-traits/src/scalar.rs:71. Neither is a CI clippy gate in this repository.

The old fused bench header said a gate fails only after all timing attempts exceed the threshold; the existing allocation check was an additional failure condition. The header now states both.

[github-actions]

Repository rules review

Repository rules review (91185c8d0498b50739507a9c49b6cf247186ac7b...23a7d10bca855f947f38a43e1b5806f5e8faf981)
Verdict: pass
Findings:
- [warn] llm-skipped (External LLM Review) <unknown>: External LLM review was skipped
  External LLM review is permanently disabled; deterministic checks only.