Fix guided generation: stop on grammar termination, include <|endoftext|>

Two fixes for the same bug (defense in depth):

1. When the grammar session terminates (isTerminated), return nil
   immediately instead of emitting the terminal token's decoded text.
   This prevents ANY special token from leaking into structured output,
   regardless of whether it's in the stop sequences list.

2. Add 'endoftext' to turnEndPatterns so <|endoftext|> (token 151643)
   is picked up from added_tokens_decoder as an additional stop token.
   Qwen3 declares eos_token as <|im_end|> (151645) but xgrammar can
   also produce <|endoftext|> as a valid grammar terminal.

Author: stikves

swift/Sources/CoreAILanguageModels/Bundle/LanguageConfig.swift

@@ -126,7 +126,7 @@ public struct LanguageConfig: Codable, Sendable, Equatable {
         // 3. Check added_tokens_decoder for turn-ending special tokens
         //    (e.g. Gemma's <end_of_turn> ID 106, Qwen's <|im_end|>)
         //    Only include tokens whose content matches known turn-ending patterns.
-        let turnEndPatterns = ["end_of_turn", "im_end", "eot_id"]
+        let turnEndPatterns = ["end_of_turn", "im_end", "eot_id", "endoftext"]
         if let addedTokens = json["added_tokens_decoder"] as? [String: Any] {
             for (idString, value) in addedTokens {
                 guard let dict = value as? [String: Any],

swift/Sources/CoreAILanguageModels/DecodingStrategies/ConstrainedDecodingStrategy.swift

@@ -324,17 +324,18 @@ extension ConstrainedDecodingStrategy.ConstrainedDecodedSequence {
                 generatedTokens.append(bestToken)
                 tokenStep += 1
 
+                if terminatedAfterAccept {
+                    finished = true
+                    return nil
+                }
+
                 let delta = ConstrainedDecodingStrategy.computeTextDelta(
                     generatedTokens: generatedTokens,
                     previousDecodedText: &previousDecodedText,
                     tokenizer: tokenizer,
                     tokenStep: tokenStep
                 )
 
-                if terminatedAfterAccept {
-                    finished = true
-                }
-
                 return GenerationResult(text: delta, tokenId: bestToken, rawLogits: logits)
             }