[AndrewSillifant]

Summary

1.7.12 adds config keys that set each Spark executor's scratch PVC. Before this, every job's scratch size was fixed by its built-in profile, so a job that outgrew its PVC could only fail. On a scale-20 AML continuous run, the scorer's 50Gi filled during the coverage aggregation's spill ("No space left on device" on both executors, 4 task failures, 3 app attempts) and the run went unscored.

Added (see CHANGELOG)

• platform.compute.spark.executor_scratch sizes every Spark job's executors, including the scorer and time-travel reads.
• bronze_executor_scratch, silver_executor_scratch and gold_executor_scratch size the stage's job in the run's mode, and win over the global key. A continuous run's bronze-verify preflight takes only the global key, the same rule as bronze_executors.
• A set size is exact: whole Gi, not scaled by cores, scale or executor count.
• One resolver (executor_scratch_size) feeds the manifest, the capacity check and scratch preflight, config_snapshot.scratch.size_per_job and the perf fingerprint, so runs with different scratch sizes do not compare as the same configuration.
• Setting a size turns platform.storage.scratch.enabled on when it is unset. An explicit enabled: false with a size is refused by commands that change data, and teardown and read commands note it.
• A size that is not whole Gi is refused by commands that change data and dropped with a note by teardown and read commands, the same as the executor memory keys.

A config that sets none of these keys requests the same PVCs as before, so results do not move.

Verification

• Unit suite: 4094 passed, including 13 new tests in tests/test_executor_scratch.py. The new tests cannot be collected without the change.
• ruff, ruff format, mypy, prose guard, and the generated docs check are clean (docs/configuration.md regenerated).
• Release gate from a clean checkout: 14 of 15 pass (version, changelog, prose, examples, ruff, format, mypy, gitleaks tree and history, cargo fmt/clippy/test, package-guard). pre-push-hook is SKIP because the throwaway checkout has no local hook.
• Spark tier not run: no file under src/lakebench/spark/scripts/ changed.
• Not yet verified on a live cluster.