testing: smoke runs a copy of the headless template, not the shipped file; extension template never loaded in Blender

[TMHSDigital]

Why it matters

The examples are smoke-gated rigorously, but the templates and snippets, the parts users copy most directly, are not. Smoke runs a copy of the headless template rather than the shipped file. Nothing keeps the two identical, so a regression in the shipped template passes CI. The extension add-on template is never loaded into Blender at all.

Evidence

• .github/workflows/blender-smoke.yml:157-164 runs tests/smoke/tmpl_gltf.py. Today it is byte-identical to templates/headless-batch-script-template/script.py (verified with diff -q), but no check enforces that.
• tests/smoke/run_smoke.py:7 claims the copies catch drift. Because nothing compares them, they hide it.
• tests/smoke/tmpl_render.py:1 is a "Reproduction of the headless-batch-scripting SKILL worked example", another hand-maintained copy.
• validate.yml:116-131 and :162-176: the 27 snippets and all template .py files only get py_compile. A snippet that calls a removed API compiles fine. See the shader-node-group inputs.new issue for a live example.
• templates/extension-addon-template/ (__init__.py, blender_manifest.toml) is never installed, registered or unregistered in any Blender run.
• templates/ai-asset-pipeline-template/: confirm whether it is smoke-run directly or also through a copy.

Suggested approach

1. Point the smoke step at templates/headless-batch-script-template/script.py directly and delete tests/smoke/tmpl_gltf.py. If a harness wrapper is needed, runpy the shipped file.
2. Add a smoke step that runs blender --command extension validate templates/extension-addon-template, then installs and enables it headless. Assert that register() adds the operator, panel and Scene pointer, and that unregister() removes them.
3. Longer term, give each snippet a minimal tests/smoke/catalog.json row, or one run_snippets.py that imports each snippet and calls its entry point on a factory-startup scene. Generalize the smoke harness to per-snippet coverage + emit a coverage table #30 closed the harness-generalisation idea; this is the missing coverage it was meant to deliver.

Done when

• No hand-copied template lives under tests/smoke/.
• The extension template is validated, registered and unregistered in CI on 5.2 and 4.5, with a falsifier: break unregister() symmetry and observe red.