perf(ci): X64 runner hosts 3x oversubscribed (intel 98.7/32 cores) while yoga is at 0.33x — rebalance worker counts #5454

Owner: infra-6a (runner hosts); measured by aprender-flow 2026-10-16 16:2xZ (read-only ssh, /proc/loadavg, nproc, ps Runner.Worker).

X64 runner hosts are unevenly loaded: intel/mac-server 32 cores, 9 workers, load1 98.7 (3.1x); framework16 16 cores, 8 workers, load1 44.9 (2.8x); yoga 22 cores, 4 workers, load1 7.3 (0.33x). gx10 (ARM64 only) 20 cores, 2 workers, load1 9.8. The #5417 merge-group shard(3) ran 45+ min on intel-clean-room-2 while yoga had ~15 idle cores.

Proposal (numbers only): intel 9->6 workers, framework16 8->6, yoga 4->6 then 8 after a re-measure. Do not unpark the 7 offline intel runners on mac-server.
Falsifiable: after step 1 intel load1 < 70 and shard(1) wall (30 min) falls over 5 MG groups; if not, thrash was not the cause.
Evidence: /mnt/nvme-raid0/cop-inbox/receipts/q-20x-1016-r11/rebalance-1640.md. Related: #5394, #5429, #5432.