fix(task-16): custom keys follow AutoCAD's simple case compare, and on COM AutoCAD routes the write

4fe2d96 matched custom keys with str.casefold(), but AutoCAD's key compare
is a simple one-to-one case compare, not Unicode full case folding.
Measured on AutoCAD 2026 (AddCustomInfo of the second spelling over the
first on a scratch document, Close(False)): Project/PROJECT, Grün/GRÜN,
é/É, ÿ/Ÿ, ł/Ł, σ/Σ, я/Я and dž/DŽ are one key ('Duplicate key'), while
Straße/STRASSE, ẞ/ß, kelvin/Kelvin-sign, µ/Μ, σς/ΣΣ and Dž/DŽ are two keys
(both stored) -- every one of those six is casefold()-equal. So with
'Straße' present, `custom={'BEFORE': 'b', 'STRASSE': 's2', 'AFTER': ...}`
was routed to SetCustomByKey, AutoCAD answered 'Key not found', and the
request raised with subject and BEFORE applied and AFTER never written --
the half-write class the fix claimed to close; a delete of 'STRASSE' over
'Straße' raised the same way; headlessly 'STRASSE' overwrote the 'Straße'
value instead of becoming a second tag; and the validator refused
{'Straße', 'STRASSE'} with a reason AutoCAD contradicts.

COM no longer routes from a Python approximation. An exact spelling goes
to SetCustomByKey; anything else is AddCustomInfo, and its 'Duplicate key'
(excepinfo scode -2145386475, measured) falls back to SetCustomByKey. A
delete is RemoveCustomByKey, and its 'Key not found' (scode -2145386476)
is reported as not deleted. Any other com_error keeps propagating. The
scode is matched before the English description because the HRESULT text
is localised.

The headless engine and the validator match on `custom_key_fold`: a
per-character fold to lowercase only when lower()/upper() are single
characters that round-trip and the character is one of the two; it never
changes string length and reproduces every measured pair (ß, ς, µ, the
Kelvin sign, Dž, ı and İ stay as they are). Docstrings in settings.py and
server.py now state that rule instead of 'case-insensitive'.

Tests: parametrized fold over the measured SAME / DISTINCT table (plus a
pin that casefold() merges the six DISTINCT pairs, which is why 4fe2d96
regressed), validator acceptance of each DISTINCT pair, headless
side-by-side storage and deletion of each pair, the fake SummaryInfo now
matches by the same fold and raises pywintypes.com_error with the
measured excepinfo tuples, the COM half-write scenario for each pair
(no SetCustomByKey guess, 'Key not found' delete tolerated, rest of the
request lands), the 'Duplicate key' fallback, and an unrelated
SummaryInfo failure still raising. The scodes were measured live before
the change; AutoCAD stopped answering COM (blocked in another agent's
command or a modal dialog) before the post-change re-run, so that run
is recorded as not done.

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

backends/ezdxf_backend.py

@@ -41,6 +41,7 @@ from .base import (
 )
 from .contracts.settings import (
     SUMMARY_FIELDS,
+    custom_key_fold,
     parse_scale,
     scale_value,
     validate_drawing_properties,
@@ -6142,24 +6143,26 @@ class EzdxfBackend(AutoCADBackend):
                     "Save as R2004 or newer first."
                 )
             custom_vars = doc.header.custom_vars
-            # AutoCAD's custom keys are case-insensitive (measured on 2026:
-            # SetCustomByKey("PROJECT") updates an existing "Project" and keeps
-            # that spelling; a second AddCustomInfo is 'Duplicate key'), and
-            # ezdxf's CustomVars are plain case-sensitive tuples, so the match
-            # is done here: an existing tag is updated in place under its
+            # AutoCAD compares custom keys with a simple per-character case
+            # compare (measured on 2026: SetCustomByKey("PROJECT") updates an
+            # existing "Project" and keeps that spelling, a second AddCustomInfo
+            # is 'Duplicate key' -- but 'Straße' and 'STRASSE' are two keys,
+            # which str.casefold() would merge), and ezdxf's CustomVars are
+            # plain case-sensitive tuples, so the match is done here with
+            # custom_key_fold: an existing tag is updated in place under its
             # stored spelling, never duplicated under a second one.
             props = custom_vars.properties
             for key, value in to_write.items():
-                folded = key.casefold()
-                hits = [i for i, (tag, _v) in enumerate(props) if tag.casefold() == folded]
+                folded = custom_key_fold(key)
+                hits = [i for i, (tag, _v) in enumerate(props) if custom_key_fold(tag) == folded]
                 if hits:
                     props[hits[0]] = (props[hits[0]][0], value)
                 else:
                     custom_vars.append(key, value)
             deleted = []
             for key in to_delete:
-                folded = key.casefold()
-                kept = [(tag, v) for tag, v in props if tag.casefold() != folded]
+                folded = custom_key_fold(key)
+                kept = [(tag, v) for tag, v in props if custom_key_fold(tag) != folded]
                 if len(kept) != len(props):
                     props[:] = kept
                     deleted.append(key)