v0.8.10: fix community feed + falsification gated behind Pyodide; point LIMITATIONS link to GitHub-rendered

loadCommunityFeed() and renderFalsificationDashboard() were called inside
enableUI(), which only runs after Pyodide finishes loading. If the Pyodide
CDN is blocked/slow/offline, enableUI() never fires and the community feed
stays stuck on "Loading..." forever despite being Pyodide-independent
(fetch + DOM only). Moved both to bootstrap before loadPyodideAndTaf().

Also repointed the in-app docs/LIMITATIONS.md link (4 langs) to the
GitHub-rendered blob URL so it renders as a page instead of raw markdown.

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

js/i18n.js

@@ -1055,7 +1055,7 @@ export const TRANSLATIONS = {
     "gamma_check.regime.swa.desc":       "γ_obs > 1.05 on random corpus = sliding-window attention signature (Mistral / Gemma family).",
     "gamma_check.regime.unknown.desc":   "Inputs out of range or γ_obs > 1 without random-corpus flag. Verify measurement.",
     "gamma_check.validity.title":          "⚠ Closed-form γ may not apply to this model",
-    "gamma_check.validity.body":           "The Padé prediction assumes natural training without explicit attention regularization. Your η falls outside the validated band [0.85, 1.15], so the closed-form is not reliable here. Trust empirical γ (from the Phase Diagram or Diagnose CLI) over the prediction. Possible causes: heavy regularization forcing near-uniform attention, fine-tuning collapse, sliding-window architecture, or non-standard training losses. See <a href=\"docs/LIMITATIONS.md\" target=\"_blank\">docs/LIMITATIONS.md</a>.",
+    "gamma_check.validity.body":           "The Padé prediction assumes natural training without explicit attention regularization. Your η falls outside the validated band [0.85, 1.15], so the closed-form is not reliable here. Trust empirical γ (from the Phase Diagram or Diagnose CLI) over the prediction. Possible causes: heavy regularization forcing near-uniform attention, fine-tuning collapse, sliding-window architecture, or non-standard training losses. See <a href=\"https://github.com/karlesmarin/tafagent/blob/main/docs/LIMITATIONS.md\" target=\"_blank\">docs/LIMITATIONS.md</a>.",
     "gamma_check.validity.fraud.hint":     "Hint: η ≪ 1 typically indicates θ marketing inflation (YaRN-style) without genuine context extension, OR attention forced near-uniform by training.",
     "gamma_check.validity.compressed.hint":"Hint: η ∈ [0.01, 0.5) is common in instruction-tuned / RLHF models where post-training has flattened the attention distribution.",
     "gamma_check.validity.overpade.hint":  "Hint: η > 1.5 may indicate an undertrained early checkpoint, a Lerch-corrected regime, or correction terms beyond the Padé approximation.",
@@ -1206,7 +1206,7 @@ export const TRANSLATIONS = {
     "gamma_check.regime.swa.desc":       "γ_obs > 1.05 sobre corpus aleatorio = firma de sliding-window attention (familias Mistral / Gemma).",
     "gamma_check.regime.unknown.desc":   "Entradas fuera de rango o γ_obs > 1 sin flag de corpus aleatorio. Verifica la medición.",
     "gamma_check.validity.title":          "⚠ La fórmula cerrada de γ puede no aplicar a este modelo",
-    "gamma_check.validity.body":           "La predicción de Padé asume entrenamiento natural sin regularización explícita de la atención. Tu η cae fuera de la banda validada [0.85, 1.15], por lo que la fórmula cerrada no es fiable aquí. Confía en γ empírico (Phase Diagram o Diagnose CLI) por encima de la predicción. Causas posibles: regularización fuerte que fuerza atención casi uniforme, colapso por fine-tuning, arquitectura sliding-window, o pérdidas no estándar. Ver <a href=\"docs/LIMITATIONS.md\" target=\"_blank\">docs/LIMITATIONS.md</a>.",
+    "gamma_check.validity.body":           "La predicción de Padé asume entrenamiento natural sin regularización explícita de la atención. Tu η cae fuera de la banda validada [0.85, 1.15], por lo que la fórmula cerrada no es fiable aquí. Confía en γ empírico (Phase Diagram o Diagnose CLI) por encima de la predicción. Causas posibles: regularización fuerte que fuerza atención casi uniforme, colapso por fine-tuning, arquitectura sliding-window, o pérdidas no estándar. Ver <a href=\"https://github.com/karlesmarin/tafagent/blob/main/docs/LIMITATIONS.md\" target=\"_blank\">docs/LIMITATIONS.md</a>.",
     "gamma_check.validity.fraud.hint":     "Pista: η ≪ 1 suele indicar inflación de θ tipo marketing (YaRN) sin extensión real de contexto, O atención forzada a casi uniforme por el entrenamiento.",
     "gamma_check.validity.compressed.hint":"Pista: η ∈ [0.01, 0.5) es habitual en modelos instruction-tuned / RLHF donde el post-entrenamiento ha aplanado la distribución de atención.",
     "gamma_check.validity.overpade.hint":  "Pista: η > 1.5 puede indicar checkpoint temprano sub-entrenado, régimen Lerch-corregido, o términos de corrección más allá de la aproximación de Padé.",
@@ -2330,7 +2330,7 @@ export const TRANSLATIONS = {
     "gamma_check.regime.swa.desc":       "γ_obs > 1.05 sur corpus aléatoire = signature de sliding-window attention (familles Mistral / Gemma).",
     "gamma_check.regime.unknown.desc":   "Entrées hors plage ou γ_obs > 1 sans flag corpus_aléatoire. Vérifiez la mesure.",
     "gamma_check.validity.title":          "⚠ La forme close de γ peut ne pas s'appliquer à ce modèle",
-    "gamma_check.validity.body":           "La prédiction de Padé suppose un entraînement naturel sans régularisation explicite de l'attention. Votre η sort de la bande validée [0.85, 1.15], donc la forme close n'est pas fiable ici. Faites confiance au γ empirique (Phase Diagram ou Diagnose CLI) plutôt qu'à la prédiction. Causes possibles : régularisation forte forçant une attention quasi uniforme, effondrement au fine-tuning, architecture sliding-window, ou pertes non standard. Voir <a href=\"docs/LIMITATIONS.md\" target=\"_blank\">docs/LIMITATIONS.md</a>.",
+    "gamma_check.validity.body":           "La prédiction de Padé suppose un entraînement naturel sans régularisation explicite de l'attention. Votre η sort de la bande validée [0.85, 1.15], donc la forme close n'est pas fiable ici. Faites confiance au γ empirique (Phase Diagram ou Diagnose CLI) plutôt qu'à la prédiction. Causes possibles : régularisation forte forçant une attention quasi uniforme, effondrement au fine-tuning, architecture sliding-window, ou pertes non standard. Voir <a href=\"https://github.com/karlesmarin/tafagent/blob/main/docs/LIMITATIONS.md\" target=\"_blank\">docs/LIMITATIONS.md</a>.",
     "gamma_check.validity.fraud.hint":     "Indice : η ≪ 1 indique typiquement une inflation marketing de θ (style YaRN) sans vraie extension de contexte, OU une attention forcée quasi uniforme par l'entraînement.",
     "gamma_check.validity.compressed.hint":"Indice : η ∈ [0.01, 0.5) est courant dans les modèles instruction-tuned / RLHF où le post-entraînement a aplati la distribution d'attention.",
     "gamma_check.validity.overpade.hint":  "Indice : η > 1.5 peut indiquer un checkpoint précoce sous-entraîné, un régime Lerch-corrigé, ou des termes de correction au-delà de l'approximation de Padé.",
@@ -3454,7 +3454,7 @@ export const TRANSLATIONS = {
     "gamma_check.regime.swa.desc":       "随机语料上 γ_obs > 1.05 = 滑动窗口注意力签名 (Mistral / Gemma 系列)。",
     "gamma_check.regime.unknown.desc":   "输入超范围或 γ_obs > 1 但未标记随机语料。请核验测量。",
     "gamma_check.validity.title":          "⚠ 闭式 γ 可能不适用于此模型",
-    "gamma_check.validity.body":           "Padé 预测假设没有显式注意力正则化的自然训练。你的 η 落在已验证带 [0.85, 1.15] 之外，因此闭式公式在此处不可靠。优先信任经验 γ (Phase Diagram 或 Diagnose CLI) 而非预测值。可能原因：强正则化迫使注意力近乎均匀、微调导致崩溃、滑动窗口架构、或非标准训练损失。详见 <a href=\"docs/LIMITATIONS.md\" target=\"_blank\">docs/LIMITATIONS.md</a>。",
+    "gamma_check.validity.body":           "Padé 预测假设没有显式注意力正则化的自然训练。你的 η 落在已验证带 [0.85, 1.15] 之外，因此闭式公式在此处不可靠。优先信任经验 γ (Phase Diagram 或 Diagnose CLI) 而非预测值。可能原因：强正则化迫使注意力近乎均匀、微调导致崩溃、滑动窗口架构、或非标准训练损失。详见 <a href=\"https://github.com/karlesmarin/tafagent/blob/main/docs/LIMITATIONS.md\" target=\"_blank\">docs/LIMITATIONS.md</a>。",
     "gamma_check.validity.fraud.hint":     "提示：η ≪ 1 通常表示 θ 营销虚标 (YaRN 风格) 而非真实上下文扩展，或训练强制注意力近乎均匀。",
     "gamma_check.validity.compressed.hint":"提示：η ∈ [0.01, 0.5) 在 instruction-tuned / RLHF 模型中常见，后训练已使注意力分布扁平化。",
     "gamma_check.validity.overpade.hint":  "提示：η > 1.5 可能表示欠训练早期 checkpoint、Lerch 修正体制、或超出 Padé 近似的修正项。",

js/main.js

@@ -159,9 +159,6 @@ function enableUI() {
   $("compare-recipe").disabled = false;
   $("compare-btn").disabled = false;
   $("inspector-btn").disabled = false;
-  // Render community feed + falsification (independent of Pyodide)
-  renderFalsificationDashboard();
-  loadCommunityFeed();
   // Restore from URL if present
   parseUrlState();
 }
@@ -4616,6 +4613,11 @@ $("longscore-example-bad-btn")?.addEventListener("click", () => {
 // Bootstrap
 // ════════════════════════════════════════════════════════════════════
 initI18n();
+// Pyodide-independent panels: render immediately so they survive a Pyodide
+// load failure (CDN blocked / offline / slow region). These use only fetch +
+// DOM, never state.pyodide — must NOT be gated behind enableUI().
+renderFalsificationDashboard();
+loadCommunityFeed();
 loadPyodideAndTaf().catch(err => {
   setStatus(`❌ Failed to initialise: ${err.message || err}`);
   console.error(err);