fix(skills): correct measured-false claims across eight skills and two rules

[TMHSDigital]

Closes #358. Closes #391.

These are corrections to confident statements in agent-facing content. Every runtime claim was re-measured, not copied from the issue.

Correction to #358 itself

#358 proposed adding a to_mesh(preserve_all_data_layers=True) warning, claiming the default drops vertex groups. Measured false: a subsurfed object with a vertex group keeps its groups in a default to_mesh() on both 4.5.11 and 5.2.1. So this PR does not add that warning.

Changes

Where	Was	Now (evidence)
always-free-bmesh rule, bmesh-load-edit-free.py, mesh-editing skill	"GC cannot free the C memory… eventually crashes"; edit-mode free() "double-frees and crashes"	Storage is freed when the wrapper dies, and tracebacks, globals, closures and modal self delay that. The finally: free() gives deterministic release, and later use raises ReferenceError (measured). Edit-mode: an ownership rule. Practice unchanged
mesh-editing	float64 foreach_set "writes garbage"	Converts correctly, but 38 vs 21 ms on 361k verts (5.2.1); 34 vs 24 ms (4.5.11)
mesh-editing	omitting do_unlink → "refuse to delete"	Already the default; remove() without the kwarg succeeded
depsgraph	each missing clear leaks and accumulates; "in some 5.x builds… crashes"	A second to_mesh() returns the same mesh (one temp mesh per evaluated object). The unsourced crash claim is removed
headless	script gets argv "after --"	Full argv (measured), sliced by the parser
headless	—	Documents --factory-startup and --python-exit-code (a traceback otherwise exits 0)
headless	export_scene.obj fails for a "file dialog"	Removed in 4.0: calling it raises AttributeError on both versions
headless	render.render "one of the few" ops in --background	Not special: mesh.primitive_cube_add also runs headless
addon-scaffolding / bl-info-migration	keep bl_info "for 4.5 fallback"; "4.0 or earlier"	Manifest alone covers 4.2+; bl_info only for ≤4.1 / legacy install
ai-mesh-cleanup, decimate_to_budget.py	calc_loop_triangles() "required, not implicit"	Lazy on both: fresh cube 12, after bmesh rewrite 48, to_mesh() of subsurf 192, all without the call. The docs note quoted was about BMesh, not Mesh
operators	REGISTER adds undo history; BLOCKING "pauses UI updates"	REGISTER = Info + redo panel; UNDO is undo; INTERNAL hides from search; BLOCKING blocks the cursor (operator flag enum docs)
ui-panels	_PT_ "enforced"/"required"	Registers; prints Warning: 'mypanel' does not contain '_PT_'… (both versions)
procedural-materials	two BSDFs to Surface "undefined"	The second link replaces the first: 1 link remains, from the newest node (both versions)
extension template README	Install from Disk, "select the directory"	extension build --source-dir . --output-dir dist (flags confirmed via --command extension build --help on 5.2.1), then install the zip; or a local repository for live edits

Evidence

Proven by live run (--factory-startup; binaries report 5.2.1 LTS and 4.5.11 LTS from .scratch/): every row above marked measured or with numbers. Results were identical on both versions except the timing figures, which are given per version.

Inspection only: the bl_options semantics (operator flag enum docs) and the bl_info version boundary (Extensions Platform shipped in 4.2).

Generated and validators:

• claude/blender-rules.md and claude/skills/blender-rules/SKILL.md are regenerated by build_claude_rules.py; --check is clean.
• check_frontmatter, check_skill_refs, check_counts and check_import_export_rules pass.

🤖 Generated with Claude Code

[TMHSDigital]

CI evidence before merge: all checks pass ( 13 pass ). Blender Smoke on PR head: Blender 5.2.2 LTS / Blender 4.5.14 LTS. Summaries (happy path; falsifiers) per leg: 141 passed, 3 skipped, 0 failed;136 passed, 6 skipped, 0 failed;143 passed, 1 skipped, 0 failed;142 passed, 0 skipped, 0 failed;