[shinaoka]

Summary

Add one short paragraph next to the existing owning-backend/linalg rule in rules/rust/performance.md:

• Explicitly prohibit repository-local scalar-accumulate kernels for contractions, GEMM, decompositions and solves; use einsum_tensors, tcontract, tlinalg, or the owning backend provider.
• State why: bypassed tuned paths, threading/SIMD policy, and numeric contracts.
• Require // contraction-kernel-ok: <reason> immediately above a genuinely unavoidable loop and disclose the exception in the PR description.

The identical paragraph is added near the top of tenferro-rs's AGENTS.md, alongside a mechanically enforced CI guard, in tensor4all/tenferro-rs#2347. This PR only adds the paragraph: no restructuring, dependencies or other rule changes.

Background: tensor4all/tensor4all-rs#1064's SimpleTensorTrain::full_tensor shipped a scalar axpy nest despite the existing owning-abstraction rule. The enforcement PR is in tenferro-rs as requested; this shared paragraph documents the same exception convention.

Validation

• Reproduced the existing validate-rules workflow's Python standard-library relative-link check: all 15 rules/index.md links resolve.
• Checked that the new paragraph is byte-identical to the paragraph in tenferro-rs AGENTS.md.
• git diff --check passed; only rules/rust/performance.md changed.
• Bounded independent GPT review covered this paragraph and the companion guard; no finding against this paragraph.
• Hosted validate-rules CI passed on head 4d1a34eee87b18f8532dea3d8ba47daeb6d70587.

Do not merge or enable auto-merge.