Keep the rolling Measurement Set working set on NVMe

[jakobtfaber]

Question

How should the active rolling Measurement Set working set live physically on ext4 under /stage/dsa110-continuum/ms instead of being read through symlinks to the slower /data filesystem, while respecting disk watermarks and preserving recoverability from raw HDF5?

Because the production-cutover map carries execution, implement the smallest safe staging change and measure it on the same 2026-01-25 window. Preserve the phase-centre invariants and approved post-validation pruning. Acceptance includes explicit free-space refusal, no unexpected-target deletion, and before/after per-tile copy and hourly-epoch wall times.

[jakobtfaber]

Implementation is live in the immediate campaign: exact target-hour + four-overlap conversion, validated atomic NVMe promotion with a 20 GiB reserve, and post-PASS pruning that retains only the two next-epoch overlap tiles. Representative 2.03 GiB copy improved from ~54 s to 2.971 s (~18x). Keeping this open until the first strict-QA epoch records the full hourly wall time.