Avoid boxing ParNew promotion-failure sizes

[skrcode]

This is a one-commit JAIPilot Cloud follow-up on the exact current head of microsoft/gctoolkit#579. It targets the original contributor branch so the change can be folded into that PR.

Change

Replace the temporary ArrayList of boxed Integers with a growable primitive int array, preserving value order and the empty-array result.

Proof

• Exact parent: 2b427ea, the current Cache duration Pattern in GenerationalHeapParser CMS remark path microsoft/gctoolkit#579 head.
• Focused tests pin zero, one, and 28 promotion-failure-size blocks and their exact order.
• The same 3 focused tests and all 124 parser tests passed before and after.
• A fixed 28-block collection fixture removed 28 boxed Integers plus ArrayList growth; five-trial isolated median moved from 733.82 ms to 537.30 ms over 5,000,000 iterations.
• The full seven-module reactor passed.

Boundary and disclosure

Regex parsing dominates real log processing, so the end-to-end gain is smaller than the isolated collection fixture. The populated field currently has no downstream reader; this is an internal allocation cleanup. Microsoft CLA handling may require folding or recreating the change under the contributor identity.

JAIPilot Cloud generated and validated the patch in skrcode/gctoolkit#2. I reviewed the exact ancestry, complete two-file diff, proof, and limitations before offering it here.