Recorded Z-projection step (moved from CellPatcher) #176

Owner decision (2026-11-06): Z projection moves from CellPatcher to morphobricks-preprocessing as an explicit, recorded preprocessing step. CellPatcher stops combining planes (CellPatcher #86, Refs #83): it accepts z_policy max/mean/sum only as a no-op for Z=1/no-Z input and refuses real Z>1 input with ZStackNotSupportedError, pointing users here.

v0.4.0 refuses Z>1 inputs (2-D planes only in v1). This step lets a user turn a Z-stack into a 2-D image deliberately, with provenance, instead of being refused.

Scope to design (owner to confirm before implementation):

• Where: a convert-style step that writes a recorded, harmonised Z=1 copy (same mechanism as the planned channel-subset copy), never an implicit projection at read time.
• Methods: max, mean, sum (and whether sum needs a dtype policy: no silent overflow or clipping; refuse or widen explicitly).
• Provenance: the source fingerprint, Z size, method, dtype handling, and the preprocessing version recorded in the output's metadata and in the run's records, so downstream bricks (CellPatcher, MorphoFeature) can see the image is a projection.
• Masks: how masks with Z relate (projected masks are not label images; likely require 2-D masks or refuse).
• Calibration: Z spacing is dropped explicitly and recorded; X/Y kept.
• Downstream: MorphoFeature run folders and CellPatcher accept the projected copy as an ordinary 2-D image; the provenance is carried.

Related: the channel-subset harmonised copy shortcut (cli_convert.py, conversion_plan.py), MorphoFeature #157 (3-D strategy).