EvolMoves: runtime extrapolation and cloud packing

runtime fit ends way before the longest family, big extrapolation gap. Need an actual packed-VM throughput test before trusting the cost math; 3 chains on one machine might just slow each other down.

runtime model: nonnegative fit using sequence count and squared alignment length, based on family-averaged local rates. The diagnostic length range is 22–1,247; applying the model at 5,105 columns is substantial extrapolation. In-sample R² is not predictive validation or calibrated cloud throughput.

files: benchmarks/output/cloud-readiness/t0-am021-57ecf2a7/wallclock_model.json, benchmarks/nute2019/scripts/fit_wallclock.py, figures/cloud-cost-notes.md, benchmarks/nute2019/cloud/IF-010-stage-resources.md.

Local runtime fit and long-family extrapolation.

Cloud-cost estimates and proposed chain packing.

packing issue in the deck: one task occupies an entire c3-standard-4 VM despite a one-CPU request. Three chains on c3-highmem-4 is an unimplemented scenario, requiring isolation, memory/thread checks and throughput trials. Historical arithmetic suggests a cost benefit only if packing does not slow work by more than about 2.17× at full occupancy.

• review questions: measure installed-runtime performance on the actual provider SKU; account for interruptions, idle VM time, storage and tree processing. Retained draws are not effective samples. The deck's historical costs are planning estimates, not current quotes or authorization to spend.