[GeophyAI]

The suite spent most of its time in about 240 tests that ran larger problems than their assertions need. These commits make them smaller and run tests/tasks in parallel on CI.

• 25 slow test files shrunk: fewer epochs, steps, shots or stages, smaller grids, and setups shared within a file. Each change keeps what the test asserts and its tolerances; for the tests shrunk most, a deliberately broken copy of the code was confirmed to fail both the old and the new test, and each file was run several times to check for flakiness. The DDP/DD comparisons start the 1-rank reference and the multi-rank run side by side.
• CI: tests/tasks runs as three shards of about equal length (tasks-plan, tasks-runner, tasks-rest); together they collect exactly the tests of tests/tasks, and a new test file lands in tasks-rest.
• tests/conftest.py: SLURM_JOB_ID is removed for the session, so on a cluster a task test no longer waits on two scontrol calls per run for metadata it does not read.
• eager_options defaults now come from sweep's EagerOptions (the eager step is compiled). The schema said use_compile: false while a task without an eager_options block ran sweep's compiled default; behaviour is unchanged, the documented default now matches it, and a test pins the two together.

Measured on the same hardware before and after (full tests/): CPU, 2 threads, cold compile cache 865 s -> 566 s; V100 1175 s -> 790 s. 0 failures on both.

🤖 Generated with Claude Code