i18n(v0.5.3): add calibration audit banner in 4 langs (EN/ES/FR/ZH)

3 new keys per language (parity 225/225/225/225):
v053.calibration.title — banner header
v053.calibration.note — full audit explanation
(SWA disabled, post_IH exploratory, GQA solid,
D_f Phase B fix)
v053.release.banner — short release tagline

UI: collapsible <details> banner inserted in TAF Card profile (after key
numbers section), auto-localized via existing applyTranslations() call.
Visible to all users regardless of locale; explains why δ_SWA is 0,
which corrections are reliable, and the D_f formula change.

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

js/i18n.js

@@ -268,6 +268,10 @@ export const TRANSLATIONS = {
     "v05.tooltip.D_14_nu_imprint": "The learned imprint slope ν = −1/(2π) times 2π equals −1. Trivial dimensional check from paper 1.",
     "v05.tooltip.D_SAGE_7":       "The central charge c=3 times |ν_imprint| times 2π equals 3. Dimensional closure connecting CFT and training imprint.",
     "v05.tooltip.nu_beta_id":     "Correlation length exponent ν times order parameter exponent β equals −1 in Phase A. Variant of D-SAGE-2.",
+
+    "v053.calibration.title":     "🔬 v0.5.3 — Calibration audit (2026-05-02)",
+    "v053.calibration.note":      "<strong>SWA correction disabled</strong> — original δ_SWA = -0.21 was fit on n=1 model (insufficient data; group-mean +0.355 with single yes-case). <strong>post_IH correction marked exploratory</strong> — group-mean ≈ 0 in re-audit (n=22 panel) does not replicate the OLS fit. <strong>GQA correction replicates</strong> (panel +0.115 vs hardcoded +0.11). <strong>D_f formula corrected for Phase B (γ&gt;1)</strong> — uses discrete cumulative sum instead of continuum approximation. LLaMA-3, Mistral, Gemma now report correct compression values.",
+    "v053.release.banner":        "🔧 v0.5.3 — Audit-driven fixes: KV compression D_f now uses discrete sum (correct for all γ); δ_SWA disabled (n=1 calibration); paper §5.2 C_V coefficient erratum (1/4 → 1/12).",
   },
 
   // ────────────────────────────────────────────────────────────────────────
@@ -320,6 +324,10 @@ export const TRANSLATIONS = {
     "v05.tooltip.D_SAGE_7":       "La carga central c=3 multiplicada por |ν_imprint| multiplicada por 2π da 3. Cierre dimensional conectando CFT con imprint de entrenamiento.",
     "v05.tooltip.nu_beta_id":     "Exponente de longitud de correlación ν multiplicado por exponente de parámetro de orden β da −1 en Fase A. Variante de D-SAGE-2.",
 
+    "v053.calibration.title":     "🔬 v0.5.3 — Auditoría de calibración (2026-05-02)",
+    "v053.calibration.note":      "<strong>Corrección SWA desactivada</strong> — δ_SWA = -0.21 original se ajustó con n=1 modelo (datos insuficientes; promedio del único caso +0.355). <strong>Corrección post_IH marcada exploratoria</strong> — promedio de grupo ≈ 0 en re-auditoría (panel n=22) no replica el ajuste OLS. <strong>Corrección GQA replica</strong> (panel +0.115 vs hardcoded +0.11). <strong>Fórmula D_f corregida para Fase B (γ&gt;1)</strong> — usa suma cumulativa discreta en lugar de aproximación continua. LLaMA-3, Mistral, Gemma ahora reportan valores correctos de compresión.",
+    "v053.release.banner":        "🔧 v0.5.3 — Correcciones por audit: D_f de compresión KV ahora usa suma discreta (correcto para todo γ); δ_SWA desactivado (calibración n=1); erratum coeficiente C_V paper §5.2 (1/4 → 1/12).",
+
     "hero.title":     "🔬 TAF Agent",
     "hero.tagline":   "Prueba <strong>CUALQUIER</strong> LLM transformer antes de gastar GPU/€.",
     "hero.subtitle":  "Todo el cómputo corre localmente en tu navegador. Gratis. Sin límites. Auditable.",
@@ -583,6 +591,10 @@ export const TRANSLATIONS = {
     "v05.tooltip.D_SAGE_7":       "La charge centrale c=3 multipliée par |ν_imprint| multipliée par 2π donne 3. Fermeture dimensionnelle reliant CFT à l'empreinte d'entraînement.",
     "v05.tooltip.nu_beta_id":     "Exposant de longueur de corrélation ν multiplié par exposant de paramètre d'ordre β donne −1 en Phase A. Variante de D-SAGE-2.",
 
+    "v053.calibration.title":     "🔬 v0.5.3 — Audit de calibrage (2026-05-02)",
+    "v053.calibration.note":      "<strong>Correction SWA désactivée</strong> — δ_SWA = -0.21 d'origine était calibrée sur n=1 modèle (données insuffisantes ; moyenne du cas unique +0.355). <strong>Correction post_IH marquée exploratoire</strong> — moyenne de groupe ≈ 0 en ré-audit (panel n=22) ne réplique pas l'ajustement OLS. <strong>Correction GQA réplique</strong> (panel +0.115 vs hardcoded +0.11). <strong>Formule D_f corrigée pour Phase B (γ&gt;1)</strong> — utilise une somme cumulative discrète au lieu d'une approximation continue. LLaMA-3, Mistral, Gemma rapportent maintenant des valeurs de compression correctes.",
+    "v053.release.banner":        "🔧 v0.5.3 — Corrections issues d'audit : D_f de compression KV utilise maintenant la somme discrète (correct pour tout γ) ; δ_SWA désactivé (calibrage n=1) ; erratum du coefficient C_V paper §5.2 (1/4 → 1/12).",
+
     "hero.title":     "🔬 TAF Agent",
     "hero.tagline":   "Testez <strong>N'IMPORTE QUEL</strong> LLM transformer avant de dépenser du GPU/€.",
     "hero.subtitle":  "Tout le calcul s'exécute localement dans votre navigateur. Gratuit. Illimité. Auditable.",
@@ -846,6 +858,10 @@ export const TRANSLATIONS = {
     "v05.tooltip.D_SAGE_7":       "中心电荷 c=3 乘以 |ν_imprint| 乘以 2π 得 3。连接 CFT 和训练印记的维度闭合。",
     "v05.tooltip.nu_beta_id":     "关联长度指数 ν 乘以序参量指数 β 在 A 相中得 −1。D-SAGE-2 的变体。",
 
+    "v053.calibration.title":     "🔬 v0.5.3 — 校准审计 (2026-05-02)",
+    "v053.calibration.note":      "<strong>SWA 修正已禁用</strong> — 原 δ_SWA = -0.21 基于 n=1 模型拟合（数据不足；唯一案例的均值为 +0.355）。<strong>post_IH 修正标记为探索性</strong> — 重审中组均值 ≈ 0（n=22 面板）未能复现 OLS 拟合。<strong>GQA 修正可复现</strong>（面板 +0.115 vs 硬编码 +0.11）。<strong>D_f 公式修正 Phase B (γ&gt;1)</strong> — 使用离散累积和代替连续近似。LLaMA-3、Mistral、Gemma 现在报告正确的压缩值。",
+    "v053.release.banner":        "🔧 v0.5.3 — 审计驱动的修复：KV 压缩 D_f 现使用离散和（适用于所有 γ）；δ_SWA 禁用（n=1 校准）；论文 §5.2 C_V 系数勘误 (1/4 → 1/12)。",
+
     "hero.title":     "🔬 TAF Agent",
     "hero.tagline":   "在花费 GPU/$ 之前，测试<strong>任意</strong> Transformer LLM。",
     "hero.subtitle":  "所有计算在您的浏览器本地运行。免费。无限制。可审计。",

js/main.js

@@ -1123,6 +1123,11 @@ function renderProfile(p, params) {
       <h3 data-i18n="tafcard.numbers_title">🔢 Key numbers (paper §26)</h3>
       <div class="taf-key-numbers">${numbersHtml}</div>
 
+      <details class="v053-audit-banner" style="margin:0.8em 0; padding:0.6em 0.8em; border:1px solid rgba(241,196,15,0.5); border-radius:6px; background:rgba(241,196,15,0.07); font-size:0.88em;">
+        <summary style="cursor:pointer; font-weight:600;" data-i18n="v053.calibration.title">🔬 v0.5.3 — Calibration audit (2026-05-02)</summary>
+        <div style="margin-top:0.5em; line-height:1.45;" data-i18n="v053.calibration.note"></div>
+      </details>
+
       <div id="whatif-container" class="whatif-box"></div>
 
       <h3 data-i18n="tafcard.fals_title">🔬 Falsification status (FALSIFICATION.md F1-F23)</h3>