The 5-level nested if/else for sorted insertion created O(2^5) MLIR
regions that crashed the CuTeDSL MLIR optimizer (SIGABRT).
New approach:
- Find-min-replace: scan 6 entries to find minimum (sequential, 1-level nesting)
- Replace the minimum if new score > min (flat conditionals by index)
- Selection sort the final 6 entries after SMEM merge (descending order)
- All conditionals are FLAT (at most 1 level of nesting)
This should avoid the MLIR optimizer explosion while producing
identical results.