fix(task-14): detach an xref out of every space, and stop the live engine inventing facts

Three blocking findings on the refs group, all of them the same shape: a
number or a refusal the caller was entitled to read as measured.

1. ezdxf xref_manage: the victim scan and the `inserts` count both walked
   `doc.modelspace()` only, then dropped the definition with
   `delete_block(safe=False)` -- the flag that suppresses the in-use check.
   An xref shown on a paper-space sheet was reported as 1 insert where the
   drawing held 2, and detaching it left that sheet INSERT with no BLOCK
   behind it: `ezdxf.recover` reports `FIX 103 UNDEFINED_BLOCK: Deleted
   entity INSERT(#34) without required BLOCK definition` and the repo's own
   drawing_audit silently deletes the sheet geometry. The scan now walks
   `doc.blocks`, which is every entity space (model space, each paper-space
   tab, every block definition) and is exactly what ezdxf's own safe check
   searches with `doc.query('INSERT[name==...]i')` -- hence the
   case-insensitive name match. `delete_block` is now `safe=True`, as the
   plan's Interfaces block pins it, so a future miss raises instead of
   writing a corrupt file.

2. COM xref_manage(action='list') reported `kind: "attach"` for every row,
   a literal constant. Re-measured on this seat's registered type library
   (C:\Program Files\Common Files\Autodesk Shared\acax25*.tlb, "AutoCAD
   2025 Type Library", installed with AutoCAD 2026): IAcadBlock's xref
   members are exactly IsXRef, Path, Name, Reload, Unload, Bind, Detach,
   XRefDatabase -- no overlay indicator exists on the interface, so an
   overlay was reported to the client as an attachment. The row now says
   `kind: None`, the same honesty the adjacent `inserts` field already had,
   and the contract plus the tool docstring say so.

3. COM xref_attach had no duplicate-block-name guard, although
   contracts/refs.py and the tool docstring both promise the refusal and
   ezdxf implements it. Two different part.dwg files, or a re-attach after a
   repath, silently re-pointed or half-wrote the definition while the
   payload claimed the new path. Added `_block_names(doc)` (Blocks.Count /
   Blocks.Item(index) / block.Name -- the members xref_manage already walks)
   and the refusal fires before AttachExternalReference is dispatched, with
   the same message the headless engine gives. Lower-cased, because AutoCAD
   symbol-table names are case-insensitive.

Tests: five new ones, each failing against the unfixed backends -- an xref
shown on Layout1 is counted and detached with the drawing, the detached file
audits clean through ezdxf.recover, delete_block's in-use check is pinned,
the live listing reports kind None (against a fake that models only the
measured members, so it cannot enshrine an invented one), and the live attach
refuses a held name (xref and plain block, either case) without dispatching.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FKFStjEFzpGyLq5bMpWxTd

server.py

@@ -8635,8 +8635,11 @@ async def xref_manage(
     `capability: "xref_live"` rather than pretended, and the refusal says which
     engine does them. Other refusals: an action outside the five, an unknown
     xref name, a 'path' action with no `new_path`. On the live engine a listed
-    row's `inserts` is null — ActiveX reports no per-xref insert count and this
-    server does not invent one.
+    row's `inserts` and `kind` are both null — ActiveX's block interface
+    carries neither a per-xref insert count nor an overlay indicator, and this
+    server does not invent them. Headlessly both are measured: `inserts`
+    counts every insert in the drawing, model space and paper space alike, and
+    `detach` removes exactly those.
     """
     await ctx.info(f"Xref {action} {name or '(all)'}")
     validated = str(validate_path(new_path)) if new_path else None