feat: sharp edge pipeline V02, tessellation presets, media cache-bust, GMSH plan

Sharp Edge Pipeline V02:
- export_step_to_gltf.py: replace BRep_Tool.Polygon3D_s (returns None in XCAF) with
  GCPnts_UniformAbscissa curve sampling at 0.3mm step — extracts 17,129 segment pairs
- Inject sharp_edge_pairs + sharp_threshold_deg into GLB extras (scenes[0].extras)
  via binary GLB JSON-chunk patching (no extra dependency)
- export_gltf.py: read schaeffler_sharp_edge_pairs from Blender scene custom props,
  apply via KD-tree to mark edges sharp=True + seam=True (OCC mm Z-up → Blender transform)
- tools/restore_sharp_marks.py: dual-pass (dihedral angle + OCC pairs), updated coordinate
  transform (X, -Z, Y) * 0.001

Tessellation:
- Admin UI: Draft / Standard / Fine preset buttons with active-state highlighting
- Default angular deflection: preview 0.5→0.1 rad, production 0.2→0.05 rad
- export_glb.py: read updated defaults from system_settings

Media / Cache:
- media/service.py: get_download_url appends ?v={file_size_bytes} cache-buster
- media/router.py: Cache-Control: no-cache for all download/thumbnail endpoints

Render pipeline:
- still_render.py / turntable_render.py: shared GPU activation + camera improvements
- render_order_line.py: global render position support
- render_thumbnail.py: updated defaults

Frontend:
- InlineCadViewer: file_size_bytes-aware URL update triggers re-fetch on regeneration
- ThreeDViewer: material panel, part selection, PBR mode improvements
- Admin.tsx: tessellation preset cards, GMSH setting dropdown
- MediaBrowser, ProductDetail, OrderDetail, Orders: various UI improvements
- New: MaterialPanel, GlobalRenderPositionsPanel, StepIndicator components
- New: renderPositions.ts API client

Plans / Docs:
- plan.md: GMSH Frontal-Delaunay tessellation plan (6 tasks)
- LEARNINGS.md: OCC Polygon3D_s None issue + GCPnts fix
- .gitignore: add backend/core (core dump from root process)

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

render-worker/scripts/still_render.py

@@ -223,8 +223,14 @@ def _import_glb(glb_file):
                        max(v.y for v in all_corners),
                        max(v.z for v in all_corners)))
         center = (mins + maxs) * 0.5
-        for p in parts:
-            p.location -= center
+        # Move root objects (parentless) to centre.  Adjusting a child's local
+        # .location by a world-space vector gives wrong results when the GLB has
+        # Empty parent nodes (OCC assembly hierarchy).  Shifting the root moves
+        # the entire hierarchy correctly.
+        all_imported = list(bpy.context.selected_objects)
+        root_objects = [o for o in all_imported if o.parent is None]
+        for obj in root_objects:
+            obj.location -= center
 
     return parts
 
@@ -244,9 +250,10 @@ def _resolve_part_name(index, part_obj, part_names_ordered):
 def _apply_material_library(parts, mat_lib_path, mat_map, part_names_ordered=None):
     """Append materials from library .blend and assign to parts via material_map.
 
-    With per-part STL import, Blender objects are named after STEP parts,
-    so matching is by name (stripping Blender .NNN suffix for duplicates).
-    Falls back to part_names_ordered index-based matching for combined-STL mode.
+    Matching priority per part:
+      1. GLB object name (strip Blender .NNN suffix + OCC _AF0/_AF1 suffix)
+      2. Prefix fallback (longest mat_map key that is a prefix of / contains part name)
+      3. Index-based via part_names_ordered (also strips _AF suffix)
 
     mat_map: {part_name_lower: material_name}
     Parts without a match keep their current material.
@@ -286,16 +293,35 @@ def _apply_material_library(parts, mat_lib_path, mat_map, part_names_ordered=Non
     # secondary: index-based via part_names_ordered (combined STL fallback)
     assigned_count = 0
     for i, part in enumerate(parts):
-        # Try name-based matching first (strip Blender .NNN suffix)
+        # 1. Name-based: strip Blender .NNN suffix, then OCC _AF0/_AF1 suffix
         base_name = _re.sub(r'\.\d{3}$', '', part.name)
+        _prev = None
+        while _prev != base_name:
+            _prev = base_name
+            base_name = _re.sub(r'_AF\d+$', '', base_name, flags=_re.IGNORECASE)
         part_key = base_name.lower().strip()
         mat_name = mat_map.get(part_key)
 
-        # Fall back to index-based matching via part_names_ordered
+        # 2. Prefix fallback: longest mat_map key that is a prefix/suffix match
+        if not mat_name:
+            for key, val in sorted(mat_map.items(), key=lambda x: len(x[0]), reverse=True):
+                if len(key) >= 5 and len(part_key) >= 5 and (
+                    part_key.startswith(key) or key.startswith(part_key)
+                ):
+                    mat_name = val
+                    break
+
+        # 3. Index-based fallback via part_names_ordered (also strips _AF suffix)
         if not mat_name and part_names_ordered and i < len(part_names_ordered):
             step_name = part_names_ordered[i]
-            part_key = step_name.lower().strip()
-            mat_name = mat_map.get(part_key)
+            step_key = step_name.lower().strip()
+            mat_name = mat_map.get(step_key)
+            if not mat_name:
+                _p2 = None
+                while _p2 != step_key:
+                    _p2 = step_key
+                    step_key = _re.sub(r'_af\d+$', '', step_key)
+                mat_name = mat_map.get(step_key)
 
         if mat_name and mat_name in appended:
             part.data.materials.clear()