feat(prompts): locale-native closing bookend to defeat English-context drift (#1118 follow-up #2)

The opening preamble from commit `47f6d69e5` works for the first
few turns, but as English context accumulates in the session (code
read into the transcript, error logs, file listings, search
results, project context), the transformer's recency bias pulls
`reasoning_content` back toward English even when the user keeps
writing in their own language. The empirical fingerprint is "model
thinks in Chinese for the first 3-4 turns, then quietly switches
to English thinking around turn 5 as more code lands in context."

Community feedback (WeChat thread on #1118 — @MuMu describes an
XML-tagged "bilingual bookend" pattern they used in another
project, and @益达 confirms the translation-accuracy problem with
fully-translated prompts) pointed at the bookend pattern: keep the
rule-heavy middle of the prompt in English (single source of
truth, model is natively multilingual), but reinforce the locale
directive at BOTH ends in native script. The opening anchors
behavior at session start; the closer sits at the maximum-
recency position right before the user's next message and
re-asserts the rule each turn.

`locale_reinforcement_closer()` returns Some for `zh-Hans` /
`zh-CN` / `zh`, `ja` / `ja-JP`, `pt-BR` / `pt`. English (and
unmatched locales) return None — system prompt stays
byte-identical to the previous behavior for English users.

The closer is appended after the previous-session handoff block
(the existing "last block" position), so it's the very last
content before the user's first message. Any future block that
needs to sit closer to the user should be added BEFORE the
closer with an updated test invariant.

Three new tests pin the contract:

* `locale_reinforcement_closer_returns_native_script_for_supported_locales`
  — each supported locale's closer is in its native script and
  explicitly mentions `reasoning_content` (the V4 knob).
* `system_prompt_bookends_zh_hans_with_preamble_and_closer` —
  the full zh-Hans system prompt contains both `## 语言要求`
  (preamble) and `## 语言再次提醒` (closer), in that order, and
  no other top-level `##` section follows the closer.
* `system_prompt_skips_locale_preamble_for_english` (extended)
  — English locale gets neither the preamble nor any of the
  three locale closers.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

crates/tui/src/prompts.rs

@@ -160,6 +160,31 @@ pub(crate) fn locale_reinforcement_preamble(locale_tag: &str) -> Option<&'static
     }
 }
 
+/// Locale-native closing reinforcement appended to the very end of the
+/// system prompt — the bookend MuMu described in the WeChat thread that
+/// prompted #1118 follow-up work.
+///
+/// The opening preamble alone is not enough: as the model accumulates
+/// English context turn-over-turn (code, error logs, search results,
+/// file listings), the recency bias of the transformer's attention
+/// drifts thinking back toward English even when the user keeps writing
+/// in their own language. A closing native-script reinforcement sits at
+/// the position closest to the user's next message — where attention
+/// weight is highest — and re-asserts the language rule right before
+/// the model generates `reasoning_content` for the turn.
+///
+/// Like the opening preamble, English (and unknown) locales return
+/// `None` and the system prompt is byte-identical to the pre-bookend
+/// behavior.
+pub(crate) fn locale_reinforcement_closer(locale_tag: &str) -> Option<&'static str> {
+    match locale_tag {
+        "zh-Hans" | "zh-CN" | "zh" => Some(LOCALE_CLOSER_ZH_HANS),
+        "ja" | "ja-JP" => Some(LOCALE_CLOSER_JA),
+        "pt-BR" | "pt" => Some(LOCALE_CLOSER_PT_BR),
+        _ => None,
+    }
+}
+
 const LOCALE_PREAMBLE_ZH_HANS: &str = "## 语言要求\n\n\
 你正在 DeepSeek TUI 中运行。无论任务上下文（代码、错误日志、文件名）\
 是英文，无论系统提示的其余部分是英文，你都必须用简体中文进行 \
@@ -194,6 +219,35 @@ Se o usuário mudar de idioma no meio da sessão, mude no próximo turno. \
 Se o usuário pedir explicitamente (por exemplo, \"think in English\"), \
 isso sobrescreve esta regra.";
 
+// ── Closing bookends (appended to the very end of the system prompt) ──
+
+const LOCALE_CLOSER_ZH_HANS: &str = "## 语言再次提醒\n\n\
+**重要：你的 `reasoning_content`（内部思考）和最终回复必须保持简体中文。** \
+无论你在这次会话中读到了多少英文代码、错误日志或文档，无论项目上下文 \
+是英文，思考过程不能漂移到英文。这是会话级硬性要求 —— 用户的语言决定 \
+你的语言，与上下文中累积的英文内容无关。除非用户明确要求切换（例如 \
+\"think in English\"），否则继续用简体中文思考和回答。";
+
+const LOCALE_CLOSER_JA: &str = "## 言語再確認\n\n\
+**重要：`reasoning_content`（内部思考）と最終的な返信は日本語で行ってください。** \
+このセッションで読み込んだ英語のコード、エラーログ、ドキュメントの量に \
+関係なく、プロジェクトコンテキストが英語であっても、思考プロセスを \
+英語に逸らさないでください。これはセッションレベルの厳格な要件であり、 \
+ユーザーの言語があなたの言語を決定します。ユーザーが明示的に切り替えを \
+要求しない限り（例：\"think in English\"）、日本語で思考し、回答し続けて \
+ください。";
+
+const LOCALE_CLOSER_PT_BR: &str = "## Reforço de Idioma\n\n\
+**Importante: seu `reasoning_content` (pensamento interno) e a resposta \
+final devem permanecer em português do Brasil.** Independentemente de \
+quanto código em inglês, logs de erro ou documentação você ler nesta \
+sessão, e independentemente de o contexto do projeto ser em inglês, o \
+processo de pensamento não pode derivar para o inglês. Este é um \
+requisito rígido em nível de sessão — o idioma do usuário define seu \
+idioma. A menos que o usuário peça explicitamente a troca (por exemplo, \
+\"think in English\"), continue pensando e respondendo em português do \
+Brasil.";
+
 /// Personality overlays — voice and tone.
 pub const CALM_PERSONALITY: &str = include_str!("prompts/personalities/calm.md");
 pub const PLAYFUL_PERSONALITY: &str = include_str!("prompts/personalities/playful.md");
@@ -553,6 +607,21 @@ pub fn system_prompt_for_mode_with_context_skills_session_and_approval(
         full_prompt = format!("{full_prompt}\n\n{handoff_block}");
     }
 
+    // 7. Locale-native closing reinforcement (#1118 follow-up #2). The
+    // opening preamble alone wasn't enough — community feedback (the
+    // WeChat thread about XML-tagged bilingual bookends) flagged that as
+    // English context accumulates turn-over-turn, the model's recency
+    // bias pulls thinking back to English. Putting the same directive at
+    // the END of the system prompt — right before the user's next
+    // message — uses recency bias *in our favor*: the model sees the
+    // native-script "keep thinking in Chinese / Japanese / Portuguese"
+    // rule immediately before it generates `reasoning_content` for the
+    // turn. English (and unknown) locales return `None` and the prompt
+    // stays byte-identical to the pre-bookend behavior.
+    if let Some(closer) = locale_reinforcement_closer(session_context.locale_tag) {
+        full_prompt = format!("{full_prompt}\n\n{closer}");
+    }
+
     SystemPrompt::Text(full_prompt)
 }
 
@@ -696,6 +765,81 @@ mod tests {
         );
     }
 
+    #[test]
+    fn locale_reinforcement_closer_returns_native_script_for_supported_locales() {
+        // English (and unknown locales) get None.
+        assert!(locale_reinforcement_closer("en").is_none());
+        assert!(locale_reinforcement_closer("fr-FR").is_none());
+        assert!(locale_reinforcement_closer("").is_none());
+
+        // Each supported locale gets a closer in its own script that
+        // explicitly tells the model "don't drift to English even as
+        // English context accumulates" — that's the load-bearing claim
+        // behind the bookend pattern.
+        let zh = locale_reinforcement_closer("zh-Hans").expect("zh closer");
+        assert!(
+            zh.contains("简体中文"),
+            "zh closer must be in Simplified Chinese"
+        );
+        assert!(
+            zh.contains("reasoning_content"),
+            "zh closer must steer reasoning_content"
+        );
+        let ja = locale_reinforcement_closer("ja").expect("ja closer");
+        assert!(ja.contains("日本語"), "ja closer must be in Japanese");
+        assert!(ja.contains("reasoning_content"));
+        let pt = locale_reinforcement_closer("pt-BR").expect("pt-BR closer");
+        assert!(pt.contains("português do Brasil"));
+        assert!(pt.contains("reasoning_content"));
+    }
+
+    #[test]
+    fn system_prompt_bookends_zh_hans_with_preamble_and_closer() {
+        // The full system prompt for zh-Hans must contain BOTH the
+        // opening preamble (`## 语言要求`) and the closing reinforcement
+        // (`## 语言再次提醒`), with the closer appearing AFTER the
+        // preamble — i.e. the prompt is "bookended" in native script,
+        // matching the empirical finding from the WeChat thread that
+        // motivated the closer.
+        let tmp = tempdir().expect("tempdir");
+        let text = match system_prompt_for_mode_with_context_skills_session_and_approval(
+            AppMode::Agent,
+            tmp.path(),
+            None,
+            None,
+            None,
+            PromptSessionContext {
+                user_memory_block: None,
+                goal_objective: None,
+                project_context_pack_enabled: false,
+                locale_tag: "zh-Hans",
+            },
+            ApprovalMode::Suggest,
+        ) {
+            SystemPrompt::Text(text) => text,
+            SystemPrompt::Blocks(_) => panic!("expected text system prompt"),
+        };
+        let preamble_pos = text
+            .find("## 语言要求")
+            .expect("zh-Hans preamble must be in prompt");
+        let closer_pos = text
+            .find("## 语言再次提醒")
+            .expect("zh-Hans closer must be in prompt");
+        assert!(
+            preamble_pos < closer_pos,
+            "closer must come after preamble (preamble={preamble_pos}, closer={closer_pos})",
+        );
+        // The closer must be the very last block — anything else after
+        // it defeats the recency-bias purpose. Skip the closer's own
+        // `## ` header before scanning.
+        let closer_header_end = closer_pos + "## 语言再次提醒".len();
+        let after_closer_body = &text[closer_header_end..];
+        assert!(
+            !after_closer_body.contains("\n## "),
+            "no other top-level section should follow the closer; got: {after_closer_body:?}",
+        );
+    }
+
     #[test]
     fn system_prompt_skips_locale_preamble_for_english() {
         // English locale → no preamble injected. Asserts the
@@ -730,6 +874,19 @@ mod tests {
             !text.contains("Requisito de Idioma"),
             "English locale must not get a pt-BR preamble: {text:?}"
         );
+        // Closer too — same bookend rule.
+        assert!(
+            !text.contains("语言再次提醒"),
+            "English locale must not get a zh closer: {text:?}"
+        );
+        assert!(
+            !text.contains("言語再確認"),
+            "English locale must not get a ja closer: {text:?}"
+        );
+        assert!(
+            !text.contains("Reforço de Idioma"),
+            "English locale must not get a pt-BR closer: {text:?}"
+        );
     }
 
     #[test]