Execute source-order folds without sorting

[PraneethMerugu]

Summary

• execute source-order folds directly over ascending physical source positions
• keep canonical folds on the existing compact-and-sort path
• specialize recurrence only on two durable traversal laws, not ordering callback identities

Evidence

• focused CPU: 80/80
• full CPU: 1,917 assertions across 47 testsets
• focused real Metal: 63/63
• full real Metal: 568/568 across 27 testsets
• launches at extents 32/300/301: 21/51/51 to 6/6/6
• warm allocations: 7,792/77,712/77,712 to 3,472/7,584/7,584 bytes
• isolated recurrence JET clean; AllocCheck clean; no LocalMath-owned load invalidations

SourceOrder retains exact dense participating ordinals for diagnostics, halt behavior, failure atomicity, empty-input behavior, and CPU/Metal parity. Canonical callable families still share the compacted recurrence MethodInstance.