Stop repeated atomic optimizer-profile writes in hot shared loops

Problem

Optimizer binary profiles performed atomic read-modify-writes for kinds already published millions of iterations earlier, forcing shared-realm workers to contend on hot per-instruction bytes.

Acceptance

• Binary operand profiling performs atomic writes only for newly observed kind bits.
• Concurrent observations converge without losing any kind.
• Normal and TSan focused coverage remains green.
• The exact eight-lane arithmetic reproducer returns near the accepted scale.
• The complete comparison publication is regenerated after the batched contention fixes.

Result

4c0d04b load-checks monotonic profile bits before atomic OR and retains race-safe convergence. The focused profile tests pass 4/4 normally and under TSan. The exact eight-lane reproducer improved from 84.0 seconds to 2.308 seconds; the July 22 full matrix and all 1,540 samples are published. The complete unit suite passes 1,404/1,404.

Completed by 4c0d04b. The hot shared arithmetic reproducer is back from 84.0 s to 2.308 s; focused normal/TSan and full-suite evidence are green.