Preserve native POINT kind and geometry in ImageJ ROI archive roundtrips #12

Native defect

Pinned/current main 0efe67f, verified against the remote. A tiny synthetic native codec diagnostic uses a PointShape at Y/X=(1.25,3.5), stable object/source metadata and Z plane index2; no biological images, pipeline execution or viewer mutation.

Observed:

• PointImageJROIShapeConverter.imagej_roi (src/polystore/roi_converters.py:547) uses ImagejRoi.frompoints without setting ROI_TYPE.POINT. The stored record is FREEHAND rather than POINT.
• load_rois_from_zip (src/polystore/roi.py) maps every non-POLYLINE record to PolygonShape. Both the production point-writer archive and an independently constructed standard ImageJ POINT archive fail with ValueError: Polygon must have at least 3 vertices, got 1.
• The native polygon writer/reader control passes exact XY and metadata checks.
• Floating XY coordinates and sidecar Z/object/source identity are present; a shape-kind failure must not be masked by a caller-side decoder.

Evidence environment: Python3.14.6, numpy2.5.1, roifile2026.2.10, own pinned PolyStore/arraybridge/metaclass-registry/zmqruntime sources. This is codec evidence only, not an installed OpenHCS/MCP/viewer or biological proof.

Required owner-level repair

Keep external ImageJ format semantics and put encode/decode behavior on the existing nominal geometry/codec declaration owners; derive dispatch from declarations/capabilities, not a parallel shape roster or extension/type ladder. Preserve POINT kind, exact YX and Z/label/object/source metadata; reject unsupported/malformed types truthfully. Protect polygon/polyline/ellipse behavior with family/native external-contract regressions.

Complete source/dependency NRA and R1 ownership evidence plus bounded native tests are required before implementation is accepted. Do not edit shared/installed dependencies or bypass the OpenHCS managed viewer.

This is a prerequisite of OpenHCSDev/openhcs#134 / draft OpenHCSDev/openhcs#153: fixing result-directory admission alone would leave point reopening broken. Parent gitlink integration follows validated dependency changes; no current scientific artifact is rewritten.