Layer 0 Axis: `compute_mode`

compute_mode decides whether execution runs serially or in parallel. It is a runtime setup choice, not a scientific treatment.

Values

Value	Status	Meaning
`serial`	operational, default	run one work unit at a time
`parallel`	operational	parallelize over a leaf-config work unit

This axis is not sweepable. Do not compare serial and parallel as model alternatives.

compute_mode: parallel requires:

Key	Values	Meaning
`parallel_unit`	`models`, `horizons`, `targets`, `oos_dates`	unit over which workers are split
`n_workers`	positive integer or `auto`	worker count

compute_mode: serial should omit parallel_unit and n_workers unless a wrapper explicitly ignores them.

`parallel_unit`	Use When
`models`	the same data/task is evaluated across model alternatives
`horizons`	the same method is evaluated across forecast horizons
`targets`	multi-target studies can run targets independently
`oos_dates`	rolling/recursive origin cells dominate runtime

Parallelism changes scheduling only. It should not change the compiled recipe, feature construction, fitted model definition, or metric definitions.

Serial default:

0_meta:
  fixed_axes:
    compute_mode: serial

Parallel over targets:

0_meta:
  fixed_axes:
    compute_mode: parallel
  leaf_config:
    parallel_unit: targets
    n_workers: 4

Parallel over out-of-sample origins:

0_meta:
  fixed_axes:
    compute_mode: parallel
  leaf_config:
    parallel_unit: oos_dates
    n_workers: auto

Invalid Pattern	Use Instead
`compute_mode: parallel_by_model`	`compute_mode: parallel`, `parallel_unit: models`
`compute_mode: parallel_by_horizon`	`compute_mode: parallel`, `parallel_unit: horizons`
`compute_mode: parallel_by_target`	`compute_mode: parallel`, `parallel_unit: targets`
`compute_mode: parallel_by_oos_date`	`compute_mode: parallel`, `parallel_unit: oos_dates`
sweeping `compute_mode`	keep fixed; sweep scientific axes only

The public registry intentionally has only serial and parallel.
Worker count belongs in leaf_config, not in the axis value.
If parallel execution fails, failure handling is still controlled by failure_policy.