Log [om-metrics] resource usage from the docker entrypoint

[aethr]

Closes #86.

wrf-run asks for 63 vCPU and 32000 MiB, and neither has been measured. This adds the same [om-metrics] start/finish lines the openmethane image emits (openmethane/openmethane#235), so one aust10km daily run is enough to size it:

just logs <log-stream-name> | grep om-metrics

Changes

• scripts/docker-entrypoint-debug.sh: a port of the openmethane wrapper, with the same fields and format. Unlike openmethane, it runs the command directly and passes on its real exit code, because setup-wrf has no entrypoint of its own.
• Dockerfile: installs tini and sets the wrapper as the ENTRYPOINT. Batch's command replaces only CMD, so wrf-run and wrf-download_geog are measured with no om-infra change.
• Interactive docker run -it sessions run the command directly and are not measured. A backgrounded command would otherwise get /dev/null as stdin.
• After forwarding a signal, the wrapper keeps waiting until the command has exited. Without this, a stopped job would end the container while run-wrf.sh was still copying partial output from /mnt/scratch to EFS.
• Tests ported from openmethane, plus checks for exit-code passthrough and the signal handling.
• docs/development.md: a short section on the metrics.

Scratch cleanup is deliberately left out of the entrypoint. run-wrf.sh already syncs and then removes its RUN_DIR on exit. A blanket rm -rf $CHK_PATH/* would risk wiping other jobs' data, since run-wrf.sh defaults CHK_PATH to the shared /mnt/scratch.

Reading the results for wrf-run

• The figures cover the whole container: WPS, the single-process real.exe and wrf.exe. mean_parallelism will therefore sit below the rank count. The time output around mpirun in run.sh covers wrf.exe on its own.
• safe_mpi_tasks() sizes ranks from os.cpu_count(). Batch uses CPU shares rather than a quota, so this is probably the host's 64 vCPUs, not the 63 requested. The running with N mpi ranks log line will confirm it.
• Ranks are capped by the host, not the domain, so a larger domain needs no more vCPU. Memory scales roughly with cell count, on top of a fixed overhead per rank. A second run on aust25km would separate the two: it also gets 64 ranks, with about 1/6 of aust10km's cells.

Testing

• make test: 38 passed.
• Built the image locally and ran it under --cpus/--memory limits. The start and finish lines parse, and the exit code passes through.
• docker stop on a container whose command has an exit trap: the trap now completes before the container exits. Before the fix it was cut off.

🤖 Generated with Claude Code

[aethr]

Closing, as the perf testing this was made for is concluded. Results are in #89 .