[Vulcan] Clarify custom nickname comma shielding limit

CHANGELOG.md

@@ -30,7 +30,7 @@ adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 - Control characters and bidi/format characters are stripped during normalization, so they can no longer survive into name fields and downstream exports.
 - Nickname delimiters containing a comma, NUL, whitespace, or control characters are ignored instead of silently corrupting the structural-comma split.
 - The token budget counts the same separator class the parser splits on, and the uniform-uppercase scan is capped, so hostile separator-heavy rows stay within the documented 65,536-token cost.
-- The structural-comma shield degrades to a bounded byte-level mask past 4 KB instead of switching off, keeping surname/given assignment stable on long rows.
+- Single-byte nickname delimiter configurations use a bounded byte-level structural-comma mask without the 4,096-byte cutoff. Configurations containing any valid multi-character or multibyte delimiter still stop shielding nickname commas above 4,096 bytes; see the README parsing limits.
 - `Smith, Jr` reads like `Smith, Junior` (Lastname plus given name) instead of a bare suffix with no given name.
 - `Parser` internals are split into `StructuralCommaSplitter`, `CommaCredentialTail`, and `SegmentParserFactory` with `TokenCredentialClass` replacing the credential classifier ints; `Name` dispatches on `::class` constants. No behavior change.
 

README.md

@@ -330,6 +330,15 @@ instead of tokenizing the source again. Explicitly supplied token arrays are
 also limited to 65,536 entries and 1,048,576 aggregate token bytes; supplying
 tokens does not bypass validation of the original string.
 
+Custom nickname delimiters have a separate comma-splitting limit. If any valid
+configured opener or closer is longer than one byte (including ASCII `<<` / `>>`
+and Unicode `«` / `»`), nickname commas are not shielded when the normalized
+input exceeds 4,096 bytes. This applies even when that longer pair is unused in
+the input. A comma inside a nickname can then become a surname/given-name
+separator and change the parsed fields. For long rows with nickname commas,
+use only single-byte delimiter pairs, such as the defaults; their byte-level
+scan has no 4,096-byte cutoff. The overall input limits still apply.
+
 ### Ambiguous inputs
 
 Some inputs have no structural signal. A comma followed only by credentials can

src/StructuralCommaSplitter.php

@@ -96,7 +96,8 @@ public static function mask(string $name, array $nicknameDelimiters): string
             return self::maskAscii($name, $pairs, $symmetric);
         }
 
-        // Bound character-array allocation for multibyte delimiters; longer input is unshielded.
+        // Any multi-character or multibyte endpoint selects this bounded character scan.
+        // Above 4096 bytes, even commas inside single-byte pairs in a mixed config are unshielded.
         if (strlen($name) > 4096) {
             return $name;
         }

tests/CommaSegmentTest.php

@@ -353,7 +353,7 @@ public function testSurnameSegmentBindsPastMidSegmentNickname(): void
 
     /**
      * Use multibyte delimiters to exercise the 4096-byte shielding limit;
-     * the ASCII scan has no length cap.
+     * the scan for exclusively single-byte delimiter pairs has no such cutoff.
      *
      * @return array<string, array{int, string, string}>
      */