[spencerkclark]

This PR adds data processing configurations and slight code additions to accommodate the output from a 36-member abrupt-4xCO2 ensemble run with CM4.

Some code / ideas are borrowed from #1565, which adds similar time-coarsening functionality.

Changes:

• Adds CM4-abrupt-4xCO2-ensemble-atmosphere-1deg-8layer.yaml and CM4-abrupt-4xCO2-ensemble-atmosphere-4deg-8layer.yaml to process the atmosphere and land data from the CM4 abrupt-4xCO2 ensemble at both 1° and 4° resolution and a Makefile rule to submit the argo jobs.
• Adds cm4-abrupt-4xCO2-ensemble-ocean-1deg-1daily.yaml, cm4-abrupt-4xCO2-ensemble-ocean-1deg-5daily.yaml, cm4-abrupt-4xCO2-ensemble-ocean-4deg-1daily.yaml, and cm4-abrupt-4xCO2-ensemble-ocean-4deg-5daily.yaml configs to process ocean and sea-ice data to daily and 5-daily and 1° and 4° resolution.
• Adds the ability to process an ensemble of runs instead of just a single run in the OM4 pipeline.
• Adds the ability to specify a time_block_mean_factor for a data stream in the OM4 pipeline to support coarsening by interval averages and labeling by the interval endpoints.

[jpdunc23]

Pre-review agent: No blockers; every code path we probed at 955619b is correct. Before merge, please add unit tests for the new time-coarsening helpers and say in the PR body whether any member is held out of the normalization stats.

Full review

Findings

#	sev	where	issue	suggested fix
1	medium	scripts/gfdl_om4/pipeline/om4/run.py: _coarsen, _select_whole_blocks, stream_output_time, the process_chunk re-key, the math.lcm read width, --num-timesteps scaling; config.py: StreamConfig.__post_init__, PipelineConfig.__post_init__	No automated test covers the new time-indexing logic. Changing _coarsen to mean(skipna=True), labelling blocks by their first timestamp, or dropping the size % factor raise all pass every repo test, and the skipna change would not fail at runtime either.	Add unit tests on a small synthetic cftime dataset. The reviewer has probe scripts that exercise each helper and can share them as test seeds.
2	medium (question)	scripts/data_process/configs/CM4-abrupt-4xCO2-ensemble-atmosphere-{1,4}deg-8layer.yaml stats	Compared with the CM4-like-AM4-random-CO2-ensemble-atmosphere-* configs, the only non-path differences are the missing stats.exclude_runs and stats.start_date. The stats therefore pool all 36 members, so a member later used for validation leaks into normalization, and the 90-day and 1-year members are weighted by length. Dropping start_date is fine for an abrupt run.	State the hold-out plan; add exclude_runs if there is one.
3	low	scripts/gfdl_om4/om4.mk dataflow_abrupt_4xco2_*_all; README "or all 36"	run.main blocks on the pipeline and run-dataflow.sh runs under set -e, so plain make launches the 36 Dataflow jobs one after another and the first failure stops the rest.	Document make -k / make -j N, or launch the jobs detached.
4	low	om4.mk dataflow_abrupt_4xco2_{1,4}deg_% vs dataflow_abrupt_4xco2_{1,4}deg_1daily_%	The comment relies on the shortest-stem rule, which arrived in GNU Make 3.82. GNU Make 3.81 (macOS /usr/bin/make) picks the first matching rule, so dataflow_abrupt_4xco2_1deg_1daily_ic_0001 would run the 5-daily config with MEMBER=1daily_ic_0001 and fail at store open.	Put the _1daily_% rules before the 5-daily ones; that order works under both versions.
5	low	configs/cm4-abrupt-4xCO2-ensemble-ocean-{1,4}deg-1daily.yaml	No start_time. The sibling cm4-piControl-1deg-1daily.yaml uses it to drop an initial-condition snapshot. If any member's store has that instant, _assert_time_alignment raises.	Confirm every member's store starts at the end of day 1.
6	low	generate_face_masks_abrupt_4xco2; face_mask_url in the ocean configs	One face-mask artifact, scanned from the stride-5 days of ic_0001, serves every member at both frequencies. A missed face raises mid-job; it does not corrupt output.	None needed. Optionally scan the 1-daily config too.
7	low	pipeline/om4/postprocess.py hfds_total_area	derivation is rebuilt from a fixed string, so the 5-daily hfds_total_area does not record that it is a 5-day block mean. The values are correct.	Carry hfds's derivation forward with append_derivation.
8	low	PR body	The body does not mention: _rotate_pairs now appends to an existing derivation instead of overwriting it; check_member_wetmask_equivalence; the shared face mask; the new smoke and dataflow targets.	Extend the description.
9	low	scripts/data_process/Makefile cm4_abrupt_4xCO2_ensemble_atmosphere	The config path is inline in the recipe, and the test_config.py regexes only check CONFIG_* variables. Changing it to a non-existent path still passes. Same pattern as the random-CO2 rule.	Optional: move it into a CONFIG_* variable.
10	nit	README config.py bullet	One line is not wrapped.	Wrap it.

Linearity assumption

Coarsening runs before rotation, regridding and hfds_total_area. This is correct only because each of those steps is linear in time with static weights (rotation angle, regrid weights, wetmask, ocean fraction). A later postprocess that is nonlinear, or that uses a time-varying field on a coarsened stream, would give mean(f(x)) != f(mean(x)) without any error. A docstring note on StreamConfig.time_block_mean_factor would guard this.

Checks run

• scripts/data_process: pytest test_config.py passes.
• New OM4 helpers probed on a synthetic 20-day cftime dataset: block means, label positions, offset re-key, partial-block and misaligned-chunk raises, NaN propagation and StreamConfig validation all behave as documented. All four ocean configs load with ic_0001 and raise without a member.

— 🤖 Drafted by Claude (Opus 5.5) via Claude Code, posted on @jpdunc23's behalf.

[jpdunc23]

LGTM.

See #1584 (comment) for some optional updates, most are very minor. It might be good to add test coverage if you have time, but non-blocking.