Address Greptile review on #669: single-source the CJK punctuation sets

The CJK terminator and closer codepoint lists were restated in the
reconciler's phrase policy and again in SentenceBoundaryClassifier's closer
walk, so adding a codepoint required parallel edits with no compiler
enforcement. The two primitive sets are now one internal Character extension
in the reconciler file, and both the phrase policy and the classifier compose
from them.

Author: FuJacob

Cotabby/Support/SentenceBoundaryClassifier.swift

@@ -39,12 +39,11 @@ enum SentenceBoundaryClassifier {
         switch text[lastIndex] {
         case "!", "?":
             return true
-        // CJK sentence terminators: the ideographic full stop, fullwidth `！` `？`, and the halfwidth
-        // ideographic stop. Unlike the ASCII period these are unambiguous (they never mark decimals,
-        // list numbers, or abbreviations), so they are terminal without classifier disambiguation.
-        // Without these a Japanese completion never registers a sentence end and generation always
-        // runs to the token budget, which is why CJK suggestions came out so long.
-        case "\u{3002}", "\u{FF01}", "\u{FF1F}", "\u{FF61}":
+        // The shared CJK terminator set (see `Character.isCJKSentenceTerminator`): unambiguous, so
+        // terminal without the period disambiguation below. Without these a Japanese completion never
+        // registers a sentence end and generation always runs to the token budget, which is why CJK
+        // suggestions came out so long.
+        case let character where character.isCJKSentenceTerminator:
             return true
         case ".":
             return isTerminalPeriod(in: text, at: lastIndex)
@@ -104,14 +103,12 @@ enum SentenceBoundaryClassifier {
 
 private extension Character {
     /// Closing punctuation that may follow a sentence terminator: straight and curly quotes,
-    /// parentheses, square brackets, and braces, plus the CJK closers (corner brackets, fullwidth
-    /// parenthesis, lenticular and angle brackets). `endsSentence` walks back past a run of these to
-    /// find the real terminator underneath, so `"done."`, `(stop!)`, and `終わり。」` register as
+    /// parentheses, square brackets, and braces, plus the shared CJK closer set (see
+    /// `Character.isCJKClosingPunctuation`). `endsSentence` walks back past a run of these to find
+    /// the real terminator underneath, so `"done."`, `(stop!)`, and `終わり。」` register as
     /// sentence ends.
     var isSentenceClosingPunctuation: Bool {
         self == "\"" || self == "'" || self == ")" || self == "]" || self == "}"
-            || self == "\u{201D}" || self == "\u{2019}"
-            || self == "\u{300D}" || self == "\u{300F}" || self == "\u{FF09}"
-            || self == "\u{3011}" || self == "\u{3009}" || self == "\u{300B}" || self == "\u{FF63}"
+            || self == "\u{201D}" || self == "\u{2019}" || isCJKClosingPunctuation
     }
 }

Cotabby/Support/SuggestionSessionReconciler.swift

@@ -562,6 +562,28 @@ private extension String {
     }
 }
 
+/// The CJK punctuation primitives, internal because they are the single source of truth shared by
+/// this file's acceptance policy and `SentenceBoundaryClassifier`'s sentence-end detection. Adding a
+/// codepoint here updates phrase boundaries, chunk binding, and the generation stop in one edit.
+extension Character {
+    /// The CJK sentence terminators: ideographic full stop `。`, fullwidth `！` `？`, and the halfwidth
+    /// ideographic stop `｡`. Unlike the ASCII period these are unambiguous (they never mark decimals,
+    /// list numbers, or abbreviations), so every consumer treats them as terminal without classifier
+    /// disambiguation.
+    var isCJKSentenceTerminator: Bool {
+        self == "\u{3002}" || self == "\u{FF01}" || self == "\u{FF1F}" || self == "\u{FF61}"
+    }
+
+    /// The CJK closing punctuation: corner brackets `」` `』` (and the halfwidth corner `｣`),
+    /// fullwidth parenthesis `）`, lenticular bracket `】`, and angle brackets `〉` `》`. Walk-backs
+    /// skip a run of these to find the real terminator underneath, and chunk binding attaches them to
+    /// the word they close.
+    var isCJKClosingPunctuation: Bool {
+        self == "\u{300D}" || self == "\u{300F}" || self == "\u{FF09}"
+            || self == "\u{3011}" || self == "\u{3009}" || self == "\u{300B}" || self == "\u{FF63}"
+    }
+}
+
 private extension Character {
     /// True when the character begins a word of a space-less script (Han, Hiragana, Katakana, Hangul,
     /// Thai, Lao, Khmer, Myanmar, ...). These scripts write words without separating spaces, so the
@@ -597,24 +619,6 @@ private extension Character {
         isLetter || isNumber
     }
 
-    /// The CJK sentence terminators: ideographic full stop `。`, fullwidth `！` `？`, and the halfwidth
-    /// ideographic stop `｡`. Declared once so `isPhraseSentenceTerminator` (phrase ends) and
-    /// `bindsToPrecedingSpacelessWord` (chunk binding) share one list instead of each restating the
-    /// four codepoints and silently drifting when one is updated.
-    var isCJKSentenceTerminator: Bool {
-        self == "\u{3002}" || self == "\u{FF01}" || self == "\u{FF1F}" || self == "\u{FF61}"
-    }
-
-    /// The CJK closing punctuation: corner brackets `」` `』` (and the halfwidth corner `｣`),
-    /// fullwidth parenthesis `）`, lenticular bracket `】`, and angle brackets `〉` `》`. Declared
-    /// once so `isPhraseClosingPunctuation` (the closer walk-back) and
-    /// `bindsToPrecedingSpacelessWord` (chunk binding) share one list instead of each restating the
-    /// codepoints.
-    var isCJKClosingPunctuation: Bool {
-        self == "\u{300D}" || self == "\u{300F}" || self == "\u{FF09}"
-            || self == "\u{3011}" || self == "\u{3009}" || self == "\u{300B}" || self == "\u{FF63}"
-    }
-
     /// The CJK opening brackets: corner brackets `「` `『` (and the halfwidth corner `｢`), fullwidth
     /// parenthesis `（`, lenticular bracket `【`, and angle brackets `〈` `《`. These lead the word
     /// they quote, so the trailing-binding rule stops before them while the punctuation-led peel