fix(examples): free hay-bale's bmesh on every path; per-run depsgraph-export OBJ

[TMHSDigital]

Closes #404. Refs #401: this fixes the hay-bale defect. The rule-vs-builder-class mismatch #401 also describes is left for the rule lint in #400.

hay-bale (#401)

• belt_depths() helper: bm = bmesh.new(), then try: … if not bm.faces: return [] … finally: pass. The early return leaked the BMesh, and the later bm.free() was outside any finally. It now frees in finally.
  • This is safe because BVHTree.FromBMesh copies the geometry, and the evidence below confirms it.
• --stray-vert falsifier path: from_mesh / to_mesh / free had no try; it is now wrapped.

depsgraph-export (#404)

• It wrote gettempdir()/depsgraph_export.obj (plus .mtl) and never removed it. A previous run's OBJ could therefore satisfy the exists and size > 0 check when an export wrote nothing.
• It now uses a pid-suffixed name and deletes .obj and .mtl before the export and in finally. A caller-supplied obj_path is still pre-deleted but kept afterwards.

Evidence

Live run (--factory-startup --python-exit-code 1; binary reports 5.2.1 LTS). Each script was run on unmodified main first, then on this branch:

Run	Before	After
hay_bale.py	exit 0	exit 0, stdout identical (every measured figure)
hay_bale.py --stray-vert (catalog falsifier)	exit 15	exit 15, stdout identical
depsgraph_export.py	exit 0; left %TEMP%\depsgraph_export.obj + .mtl behind	exit 0; stdout differs only in the temp file name and export ms; no file left behind
depsgraph_export.py --unevaluated (falsifier)	n/a	exit 5, unchanged contract

check_exit_code_readme and check_helper_drift 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: 143 passed, 1 skipped, 0 failed;142 passed, 0 skipped, 0 failed;141 passed, 3 skipped, 0 failed;136 passed, 6 skipped, 0 failed;