# Rule notes

The planner treats `lower` and `upper` as offsets from `nominal`. The resolution target is half of the bilateral tolerance width.

- accuracy at or below the target selects five samples;
- accuracy within three times the target selects three samples;
- wider tolerances select one sample;
- cylinders use a scan strategy with at least five points;
- repeated-point uncertainty is divided by the square root of sample count;
- scan uncertainty receives an additional conservative 0.8 factor;
- a device must support the feature kind and reach the absolute nominal value;
- ranked candidates prefer lower accuracy score, then a stable device identifier.

An empty feature list has full planning coverage by definition. An unmeasurable feature produces a warning, reduces capability, and does not create a fabricated step. Observation classification is inclusive at both tolerance boundaries; missing observations are `Unmeasured`.

Statistics use population standard deviation for deterministic batch summaries. Capability indices use the supplied uncertainty as the process spread proxy and report a conservative adequacy flag. These are planning and review metrics, not a laboratory uncertainty budget.

Execution ordering places datum-aligned operations first and can then group steps by device. Batch size must be positive to create batches; non-positive limits intentionally produce no batches. Release gates require complete plan coverage, capable devices, complete observations, no unmeasured verdicts, passing evidence, and no unresolved plan warnings.