Certify two-process tile concurrency on H17

[jakobtfaber]

Question

Can two isolated tile worker processes improve hourly-epoch throughput on H17 without CASA signal/thread failures, storage contention, CPU oversubscription, GPU contention, nondeterministic calibration, or degraded images?

Because the production-cutover map carries execution, prototype the minimum bounded process-level concurrency after fast staging and Stage-0 decisions are settled. Partition WSClean threads explicitly and compare the same 2026-01-25 checkpoint for wall time, resource peaks, flag-mask equivalence, image MAD, catalog recovery, and all strict QA gates. Keep sequential execution if the evidence does not show a safe improvement.

[jakobtfaber]

Two-process mode is implemented and certified on the same two MS inputs: 300.163 s sequential vs 175.024 s parallel (-41.7%), 1.500 vs 2.401 GiB peak RSS, identical FLAG hashes/fractions, and image MAD/peak differences below 0.0002%. Production runner now uses two isolated workers (one GPU, 20 WSClean threads each). Keeping open until the immediate campaign produces the first full strict-QA mosaic.