[shinaoka]

The CPU placement conventions settled in tensor4all-meta become audit items.

• rules/rust/performance.md: "Threading Principles" becomes "CPU Parallelism".
  • It links the design as the source of truth: core-boundary-and-execution-ownership.md "Where placement is decided" (rules 1 to 13, "faer and spindle"), plus cpu-placement-token.md and spindle.md. These come from design: one splittable placement token for CPU fan-out tensor4all-meta#41, #43, #44 and #45.
  • Ten checkable items follow, most with the tool that checks them:
    • fan-out only through the token (cargo deny, clippy disallowed-methods);
    • no ambient pool (global_pool_unstarted() test);
    • decide before entering a pool;
    • budgets only shrink;
    • one source of truth for policy;
    • vendor libraries through Exec::external;
    • spindle confined to t4a-faer, with the tlinalg batch shapes (cargo deny, wrappers = t4a-faer);
    • no Rayon type in public signatures (cargo public-api);
    • barrier-bearing kernels use broadcast;
    • concurrency claims backed by the ledger.
  • Until a repository runs on strided_exec::Exec, an audit reports each departure as a migration gap rather than a defect.
• The four previous threading bullets are merged into these items (single source of thresholds, serial short-circuit, no ambient pool, provider-owned BLAS/OpenMP teams).
• rules/common/performance.md, rules/index.md and the tensor4all-rules-audit skill point to the section.

Docs only.

🤖 Generated with Claude Code