fix(gmsh,med): restore group metadata when converting MED to MSH

[fznoussi]

When converting a native GMSH-generated MED file to MSH format:

$ meshio convert mesh.med output.msh

the output contained a single undifferentiated block with no metadata:

$ meshio info output.msh

-> 1 block of 264 triangles, no groups, no physical names

The root cause was that the MED format stores element families as integer IDs mapped to group names:

cell_tags = [-6, -5, -4, -3, -2, -1, ...] # one per cell
cell_tags_mapping = {-1: ['subgroup'], -2: ['subgroup'], ...}

while the GMSH MSH writer requires a completely different metadata structure to produce a valid .msh file:

gmsh:physical # physical tag per cell
gmsh:geometrical # entity tag per cell
gmsh:dim_tags # (dim, entity_tag) per node
field_data # {"subgroup": [physical_tag, dimension]}

_convert_med_tags_to_gmsh() was introduced to bridge the two formats but its initial implementation only renamed cell_tags to gmsh:physical without splitting blocks, reconstructing entity tags, or rebuilding field_data. As a result the MSH writer had no information to populate $PhysicalNames, $Entities or $Nodes sections correctly.

Fix - _convert_med_tags_to_gmsh(): full metadata reconstruction

The function was completed in five steps.

Step 1 - Build physical tag mapping from MED family groups

family_groups = {-1: ['subgroup'], -2: ['subgroup'], ...}

Unique group names are extracted and assigned sequential physical
tags starting from 1:

group_to_phys = {'subgroup': 1}

Each family ID is then mapped to its physical tag:

fam_to_phys = {-1: 1, -2: 1, -3: 1, -4: 1, -5: 1, -6: 1}

Step 2 - Choose split criterion

Two cases are handled:

• Round-trip (GMSH -> MED -> MSH): gmsh:geometrical is already present in cell_data -> split by gmsh:geometrical
• Native MED (GMSH .med export -> MSH): only cell_tags exist -> split by cell_tags (MED family IDs)

Step 3 - Split blocks and generate per-cell tags

For each unique tag value within each cell block, a boolean mask
selects the matching cells. A new CellBlock is created for each
subset and assigned a sequential entity tag (entity_counter).
The family ID is translated to a physical tag via fam_to_phys.

Result for the test mesh:
before: 1 block x 264 triangles, tags mixed after: 6 blocks x 44 triangles, one per geometric entity

Step 4 - Generate gmsh:dim_tags for nodes

For each node, the topological dimension (cb.dim) and the entity
tag of the first block that references it are recorded in a
(n_points, 2) array. Nodes shared between entities are assigned
to the first entity encountered, which is consistent with the
GMSH convention.

Step 5 - Reconstruct field_data

field_data is rebuilt in the format expected by the GMSH writer:

{'subgroup': [1, 2]}   # name -> [physical_tag, dimension]

Addition - _cleanup_med_fields(): fallback for single-tag blocks

When no split is needed (only one unique tag value per block), a lightweight cleanup function handles the conversion instead:

• Renames cell_tags to gmsh:physical if not already present
• Removes point_tags from point_data
• Strips MED-specific keys (med:*) and malformed field_data entries that the GMSH writer does not understand

Imports added to gmsh/main.py

import numpy as np
from .._mesh import CellBlock, Mesh

Confirmed against mesh.med generated by GMSH 4.13:

Geometry : Box with BooleanDifference sphere cutout
Groups : Physical Surface("subgroup")
Cell types: triangles (6 families) + tetra (1 family)

Before: output.msh -> 1 block, no $PhysicalNames, no $Entities
After : output.msh -> 6 blocks, $PhysicalNames contains "subgroup",
$Entities and $Nodes sections correctly populated

Fixes nschloe#1541