fix(templates): keep every mesh, handle shared and parented imports, bake modifier stacks in order

[TMHSDigital]

Closes #467
Closes #468
Closes #462

What changed

#467: templates/ai-asset-pipeline-template/pipeline.py

• Every mesh is kept. The template no longer picks max(meshes, key=tris). It gives shared mesh data one copy per object, unparents keeping the world matrix, applies rotation and scale in one transform_apply, then joins every part with object.join under temp_override. The README states this behavior.
• The dead unit check is gone. read_factory_settings always leaves the scene metric at scale 1.0, so the old check could never fire. A geometry check replaces it: the largest world extent must be within 0.01 m to 100 m (× scale_length), or the script exits 8.
• LOD budgets are asserted. An LOD whose evaluated triangle count is over its budget exits 9.
• A RuntimeError from transform_apply or join exits 12. I skipped 10 and 11 because the shared render gates use them, and the README says so.
• The redundant apply_selected_mesh_transforms() call inside export_preset is removed: every part is already applied by then. This also removes the object-op-in-a-loop pattern from the export loop.

#468: templates/headless-batch-script-template/

• bpy.ops.object.modifier_apply per object in a loop is replaced by a stack bake. bpy.data.meshes.new_from_object(obj.evaluated_get(dg), preserve_all_data_layers=True, depsgraph=dg) bakes the whole stack in order, with one depsgraph evaluation for all objects.
• The README now matches this: the new modifier runs last and the whole stack is applied. The gotcha section explains why the operator path reverses the order, with measured numbers.
• tests/check_example_rules.py now scans templates/ for the existing two rules.
• It also adds a templates-only prefer-data-over-ops-in-loops check. A for loop that reaches bpy.ops.object.* fails the check, whether the call is direct or goes through same-file functions. An # ops-loop-exempt: <why> comment on the loop exempts it.
• This lives in its own section and functions (template_scripts, check_ops_in_loops), so the parallel bulk-write lint work in examples: png-exr-alpha teaches scene image_settings for Image.save(); missing falsifier and per-element writes elsewhere #474 rebases cleanly. Examples are not held to this check, because some loop per part over operators that have no data-API equivalent (uv.smart_project in lightmap-uv-channel).

#462: shared and parented imports

• snippets/export_preset_{unity,godot,unreal}.py and the prelude in skills/engine-export-presets/SKILL.md now do two things before the single transform_apply: copy multi-user data per object, and unparent keeping the world matrix. The Unreal snippet is now exactly 75 lines, the documented maximum.
• skills/ai-mesh-cleanup/SKILL.md gains isolate_mesh_data and clear_parent_keep_transform, and rot_scale_is_identity now reads matrix_world. The skill also notes that origin_to_base needs single-user data, that a scene-units check cannot catch centimeters written as meters (measure the geometry instead), and adds common mistake fix: remove stray context-mode block from CLAUDE.md and substantiate verification claims #8.
• transform_apply(isolate_users=True) exists on both 4.5.11 and 5.2.1 (measured). The skill mentions it but uses the data copy instead, because origin_to_base also needs single-user data, not just the apply.
• In examples/export-preset-axis, only a comment changed: it said "same body as the snippet", which is no longer true.

Smoke, in blender-smoke.yml (both matrix series)

• tests/smoke/make_pipeline_glb.py now takes --parented and --radius. The parented fixture is a Root empty rotated 90° on X, scaled 2x, at z=3. Under it are a Body box and two wheels that share one mesh (users == 2) and have different rotations.
• New tests/smoke/check_pipeline_glb.py re-imports the source and lod0.glb. Exit codes: 3 when lod0.glb has more than one mesh object, 4 when its world bbox differs from the source's, 5 when the origin is not on the world minimum Z, 6 when the node has rotation or scale.
• tests/smoke/make_input.py also writes stack.blend, a cube with a live SUBSURF at levels 1. New tests/smoke/check_glb_tris.py asserts the GLB's triangle count.
• New CI steps:
  • The headless template runs on stack.blend with --apply-modifier TRIANGULATE, and the GLB must have 48 tris.
  • The pipeline runs on the parented fixture, and its output must pass check_pipeline_glb.py.
  • A falsifier runs the pipeline on a radius-1000 sphere and passes only on exit 8.

Evidence

Live run (Blender 5.2.1 LTS .scratch\blender-5.2.1-windows-x64\blender.exe and 4.5.11 LTS .scratch\blender-4.5.11-windows-x64\blender.exe; both binaries report those versions, and the two gave identical results)

Case	Result
headless template, stack.blend + TRIANGULATE	exit 0; 26 verts / 48 faces; check_glb_tris 48 = 48, exit 0
headless template, input.blend + SUBSURF (existing CI case)	exit 0
headless template, empty.blend	exit 2 (unchanged)
pipeline, default sphere 1024,256,64	exit 0; LOD tris 1024 / 256 / 64
pipeline, parented shared-mesh fixture	exit 0; "Joined 3 part(s)"; check_pipeline_glb exit 0, bbox [-2.2179,-1.0,2.0,2.0,1.2,4.0] matches the source, origin z = 2.0000 = min z
pipeline, radius-1000 sphere	exit 8 ("largest extent 2000.0014 m is outside [0.01, 100.0] m")

Falsification probes (live, both versions, scratch copies of the script with one step removed)

Probe	Measured
isolate_mesh_data removed, parented fixture	exit 12: Cannot apply to a multi user: Object "WheelL", Mesh "WheelMesh"
clear_parent_keep_transform removed	pipeline exit 0, check_pipeline_glb exit 5: origin z = 3.0000 against min z = 2.0000
join_meshes replaced by the old max(meshes, key=tris)	pipeline exit 0, check_pipeline_glb exit 4: bbox drift 3.8179 (only WheelL shipped)
decimate_to_budget removed, sphere	exit 9: Fixture_LOD0 has 2208 triangles, over its budget of 1024
old pipeline.py from main, parented fixture	exit 1, uncaught RuntimeError: Cannot apply to a multi user
old script.py from main, stack.blend + TRIANGULATE	check_glb_tris exit 3: 72 triangles against 48 expected (4.5.11 also printed Info: Applied modifier was not first)
check_ops_in_loops on the old script.py and the old pipeline.py	flags script.py:84 (add_and_apply_modifier() -> bpy.ops.object.modifier_apply) and pipeline.py:376 (export_preset() -> bpy.ops.object.transform_apply); the current tree is clean

Two things have no CI falsifier:

• The LOD-budget check (exit 9). A budget of 1 turned out to be reachable: collapse decimated the sphere to exactly 1 triangle on both versions, so that is not a natural failing input. The check's falsifier is the local probe above.
• The matrix_world identity check. Switching it back to matrix_basis alone fails nothing, because the unparent step already makes the two matrices equal. The matrix_world read protects the skill's pattern when someone keeps the parents.

Inspection only

• The snippet and skill copies of apply_selected_mesh_transforms are the same code as the template path proven live above. I did not run them separately.
• The new CI steps were checked against the local commands they mirror. They are proven in CI on this PR, not locally on Linux.

Local validators

python tests/run_all.py: 28 passed. The 3 failures were test_bump_kind, test_plugin_content_changed and test_release_gate. They fail only under PowerShell, where bash is missing (exit 127), and pass under Git Bash. tests/check_example_rules.py: 145 scripts plus 3 templates, clean.

🤖 Generated with Claude Code