fix(i18n): cover modes section + form labels + tooltips that were missed

The previous commit added data-i18n to most elements but missed:
- modes.title h2 (the '🎯 Mode' header)
- modes.tip tooltip (the 'Four ways to use the tool' explanation)
- All 4 mode tabs (Profile/Compare/Ask/Recipe buttons)
- modes.desc paragraph
- All form field labels in Profile mode (theta, T_train, T_eval, n_attn,
n_kv, d_head, n_layers, n_params, has_swa)
- All Profile form tooltips (param.theta.tip etc.)
- profile.tip and compare.tip tooltips
- Yes/No options in has_swa select

Plus added FULL help modal translations for all 4 languages (EN/ES/FR/ZH):
- help.title, help.what.title/body, help.modes.title/profile/compare/ask/recipe
- help.recipes.title, help.add_models.title/preset/hf/manual
- help.audit.title/body, help.synthesis.title/body
- help.params.title, help.verdicts.title + verdict.yes/deg/no
- help.privacy.title/body, help.source.title/body

Now switching to ES/FR/ZH actually translates everything visible.

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

index.html

@@ -151,19 +151,21 @@
 
     <!-- Mode toggle -->
     <section id="mode-section">
-      <h2>🎯 Mode <span class="info"><span class="tooltip"><strong>Four ways to use the tool</strong>.<br>
+      <h2><span data-i18n="modes.title">🎯 Mode</span>
+        <span class="info"><span class="tooltip" data-i18n="modes.tip"><strong>Four ways to use the tool</strong>.<br>
         <strong>📇 Profile</strong>: paste a model id → all 5 recipes at once = TAF Card.<br>
         <strong>🆚 Compare</strong>: 2-3 models side-by-side on one recipe.<br>
         <strong>💬 Ask</strong>: free-form question, browser LLM picks the recipe.<br>
         <strong>📋 Recipe</strong>: manual selection with full form control.
-        </span></span></h2>
+        </span></span>
+      </h2>
       <div class="mode-tabs">
-        <button class="mode-btn active" data-mode="profile">📇 Profile a model</button>
-        <button class="mode-btn" data-mode="compare">🆚 Compare models</button>
-        <button class="mode-btn" data-mode="ask">💬 Ask plain English</button>
-        <button class="mode-btn" data-mode="recipe">📋 Pick recipe</button>
+        <button class="mode-btn active" data-mode="profile" data-i18n="modes.profile">📇 Profile a model</button>
+        <button class="mode-btn" data-mode="compare" data-i18n="modes.compare">🆚 Compare models</button>
+        <button class="mode-btn" data-mode="ask" data-i18n="modes.ask">💬 Ask plain English</button>
+        <button class="mode-btn" data-mode="recipe" data-i18n="modes.recipe">📋 Pick recipe</button>
       </div>
-      <p id="mode-desc" class="recipe-desc">
+      <p id="mode-desc" class="recipe-desc" data-i18n="modes.desc">
         <strong>Quickest start</strong>: paste any HuggingFace model id (e.g. <code>meta-llama/Meta-Llama-3-8B</code>),
         click Profile. See all 5 recipes scored in seconds.
       </p>
@@ -175,7 +177,7 @@
         💡 Quick start: pick any preset → click Generate. Or paste a model id from <a href='https://huggingface.co/models?library=transformers&sort=trending' target='_blank'>HF Hub trending</a> → 📥 Fetch → Generate.
       </div>
       <h2><span data-i18n="profile.title">📇 Profile a model</span>
-        <span class="info"><span class="tooltip">
+        <span class="info"><span class="tooltip" data-i18n="profile.tip">
           <strong>One-click full diagnosis</strong>. Paste any HF model id (or pick preset).
           Tool runs all 5 recipes (long-context, KV-compression, custom-vs-API, budget,
           hardware) and produces a single <strong>TAF Card</strong> showing verdict per
@@ -207,42 +209,42 @@
 
       <div class="form-grid" id="profile-form">
         <div class="form-field">
-          <label>θ (rope_theta) <span class="info"><span class="tooltip">RoPE base frequency from <code>config.rope_theta</code>.</span></span></label>
+          <label><span data-i18n="param.theta">θ (rope_theta)</span> <span class="info"><span class="tooltip" data-i18n="param.theta.tip">RoPE base frequency from <code>config.rope_theta</code>.</span></span></label>
           <input type="number" id="profile-theta" value="500000" />
         </div>
         <div class="form-field">
-          <label>T_train <span class="info"><span class="tooltip">Max training context. From <code>max_position_embeddings</code>.</span></span></label>
+          <label><span data-i18n="param.T_train">T_train</span> <span class="info"><span class="tooltip" data-i18n="param.T_train.tip">Max training context. From <code>max_position_embeddings</code>.</span></span></label>
           <input type="number" id="profile-T_train" value="8192" />
         </div>
         <div class="form-field">
-          <label>T_eval (your target) <span class="info"><span class="tooltip">Inference context length you'll actually serve. The key knob.</span></span></label>
+          <label><span data-i18n="param.T_eval">T_eval (your target)</span> <span class="info"><span class="tooltip" data-i18n="param.T_eval.tip">Inference context length you'll actually serve. The key knob.</span></span></label>
           <input type="number" id="profile-T_eval" value="32000" />
         </div>
         <div class="form-field">
-          <label>n_attention_heads</label>
+          <label data-i18n="param.n_attn">n_attention_heads</label>
           <input type="number" id="profile-n_attn" value="32" />
         </div>
         <div class="form-field">
-          <label>n_kv_heads</label>
+          <label data-i18n="param.n_kv">n_kv_heads</label>
           <input type="number" id="profile-n_kv" value="8" />
         </div>
         <div class="form-field">
-          <label>head_dim</label>
+          <label data-i18n="param.d_head">head_dim</label>
           <input type="number" id="profile-d_head" value="128" />
         </div>
         <div class="form-field">
-          <label>n_layers</label>
+          <label data-i18n="param.n_layers">n_layers</label>
           <input type="number" id="profile-n_layers" value="32" />
         </div>
         <div class="form-field">
-          <label>n_params (e.g. 8e9)</label>
+          <label data-i18n="param.n_params">n_params (e.g. 8e9)</label>
           <input type="text" id="profile-n_params" value="8e9" />
         </div>
         <div class="form-field">
-          <label>Has SWA?</label>
+          <label data-i18n="param.has_swa">Has SWA?</label>
           <select id="profile-has_swa">
-            <option value="false" selected>No</option>
-            <option value="true">Yes</option>
+            <option value="false" selected data-i18n="common.no">No</option>
+            <option value="true" data-i18n="common.yes">Yes</option>
           </select>
         </div>
       </div>
@@ -256,7 +258,7 @@
         💡 Try: paste 3 popular 7-8B models (Meta-Llama-3-8B, Mistral-7B-v0.1, Qwen/Qwen2.5-7B), pick recipe X-2, T_eval=16000. See which best handles long context.
       </div>
       <h2><span data-i18n="compare.title">🆚 Compare models side-by-side</span>
-        <span class="info"><span class="tooltip">
+        <span class="info"><span class="tooltip" data-i18n="compare.tip">
           <strong>Same recipe, multiple models</strong>. Pick 2-3 candidate models and
           one recipe. See verdicts in a single comparison table.<br><br>
           <strong>Use case</strong>: "I need long-context retrieval at 16K — which is

js/i18n.js

@@ -73,6 +73,55 @@ export const TRANSLATIONS = {
     "profile.hf_placeholder":  "e.g. meta-llama/Meta-Llama-3-8B or Qwen/Qwen2.5-7B",
     "compare.hf_placeholder":  "HF model id (e.g. meta-llama/Meta-Llama-3-8B)",
 
+    // Form parameters
+    "param.theta":         "θ (rope_theta)",
+    "param.theta.tip":     "<strong>RoPE base frequency</strong> from <code>config.rope_theta</code>. Higher = more long-range capacity.",
+    "param.T_train":       "T_train",
+    "param.T_train.tip":   "<strong>Max training context</strong>. From <code>max_position_embeddings</code>. Beyond this is extrapolation.",
+    "param.T_eval":        "T_eval (your target)",
+    "param.T_eval.tip":    "<strong>Your target inference context</strong>. The whole question is: will the model behave well at THIS length?",
+    "param.n_attn":        "num_attention_heads",
+    "param.n_kv":          "num_key_value_heads",
+    "param.d_head":        "head_dim",
+    "param.n_layers":      "num_hidden_layers",
+    "param.n_params":      "n_params (e.g. 8e9)",
+    "param.has_swa":       "Has SWA?",
+    "common.yes":          "Yes",
+    "common.no":           "No",
+
+    // Mode tooltips
+    "modes.tip":           "<strong>Four ways to use the tool</strong>.<br><strong>📇 Profile</strong>: paste a model id → all 5 recipes at once = TAF Card.<br><strong>🆚 Compare</strong>: 2-3 models side-by-side on one recipe.<br><strong>💬 Ask</strong>: free-form question, browser LLM picks the recipe.<br><strong>📋 Recipe</strong>: manual selection with full form control.",
+    "profile.tip":         "<strong>One-click full diagnosis</strong>. Paste any HF model id (or pick preset). Tool runs all 5 recipes (long-context, KV-compression, custom-vs-API, budget, hardware) and produces a single <strong>TAF Card</strong> with verdict per dimension + key numbers + architecture classification.<br><br><strong>Use case</strong>: \"I'm evaluating Qwen2.5-32B for production — what's its full viability profile?\" → paste id → Profile → done.",
+    "compare.tip":         "<strong>Same recipe, multiple models</strong>. Pick 2-3 candidate models and one recipe. See verdicts in a single comparison table.<br><br><strong>Use case</strong>: \"I need long-context retrieval at 16K — which is best: Llama-3-8B, Mistral-7B, or Qwen-7B?\" → pick 3 + X-2 + 16K → see winner.",
+
+    // Help modal
+    "help.title":               "📘 TAF Agent — User Manual",
+    "help.what.title":          "What does it do?",
+    "help.what.body":           "Predicts <strong>practical viability</strong> of any transformer LLM <em>before you spend GPU/$</em>. Answers questions like \"will this model work at L=32K?\" or \"should I train custom or use API?\" using deterministic Python formulas (TAF — Thermodynamic Attention Framework).",
+    "help.modes.title":         "How to use — 4 modes",
+    "help.modes.profile":       "<strong>📇 Profile</strong>: paste model id → all recipes at once = TAF Card. <strong>Best starting point</strong>.",
+    "help.modes.compare":       "<strong>🆚 Compare</strong>: 2-3 models side-by-side on same recipe. Best when choosing between candidates.",
+    "help.modes.ask":           "<strong>💬 Ask plain English</strong>: free-form question, in-browser LLM picks the recipe. Best for casual exploration.",
+    "help.modes.recipe":        "<strong>📋 Recipe + form</strong>: manual selection, full parameter control. Best when you want exact control.",
+    "help.recipes.title":       "The 5 recipes available",
+    "help.add_models.title":    "Adding new models (3 ways)",
+    "help.add_models.preset":   "<strong>Preset list</strong>: 11 popular models curated. Just select from dropdown.",
+    "help.add_models.hf":       "<strong>HF Hub fetch</strong>: paste any model id (e.g. <code>Qwen/Qwen2.5-32B-Instruct</code>), click 📥 Fetch. Browser downloads <code>config.json</code> directly from HuggingFace, fills the form. Works for any public model.",
+    "help.add_models.manual":   "<strong>Manual</strong>: fill the form fields directly with values from the model card.",
+    "help.audit.title":         "The audit chain",
+    "help.audit.body":          "Every result shows the full <strong>Computation Chain</strong> — each formula step with its inputs, output, and interpretation. Click any step to expand. Cite section numbers (§26.1, §19.1, etc.) refer to the underlying paper for derivation.",
+    "help.synthesis.title":     "The plain-English answer",
+    "help.synthesis.body":      "After the deterministic chain runs, an in-browser LLM (Qwen2.5-0.5B, ~350MB cached after first load) synthesizes a plain-English summary. The numbers above are <em>always correct</em> (deterministic Python); the synthesis is LLM-generated — verify against the chain if in doubt.",
+    "help.params.title":        "Common parameters explained",
+    "help.verdicts.title":      "What to look for in verdicts",
+    "help.verdict.yes":         "<strong style=\"color:#3fb950;\">YES / GO</strong> — proceed with confidence; numbers support the choice.",
+    "help.verdict.deg":         "<strong style=\"color:#d29922;\">DEGRADED / TINY-MODEL</strong> — works but with caveats; read the action.",
+    "help.verdict.no":          "<strong style=\"color:#f85149;\">NO / MEMORY-LIMITED</strong> — don't proceed as-is; mitigation provided.",
+    "help.privacy.title":       "Privacy",
+    "help.privacy.body":        "Everything runs in your browser. No telemetry, no analytics, no data sent anywhere. Even the LLM model runs locally via WebGPU/WebAssembly. Your model_ids and questions never leave this page.",
+    "help.source.title":        "Source & paper",
+    "help.source.body":         "Source code: <a href=\"https://github.com/karlesmarin/tafagent\" target=\"_blank\">github.com/karlesmarin/tafagent</a><br>Paper: <em>Marin 2026 — Transformer Thermodynamics</em> (arXiv forthcoming)",
+
     "footer.text":             "© 2026 Carles Marin · Apache-2.0 · independent research · the tool that closes the loop of the paper.",
   },
 
@@ -139,6 +188,55 @@ export const TRANSLATIONS = {
     "profile.hf_placeholder":  "ej. meta-llama/Meta-Llama-3-8B o Qwen/Qwen2.5-7B",
     "compare.hf_placeholder":  "ID modelo HF (ej. meta-llama/Meta-Llama-3-8B)",
 
+    // Parámetros del formulario
+    "param.theta":         "θ (rope_theta)",
+    "param.theta.tip":     "<strong>Frecuencia base RoPE</strong> de <code>config.rope_theta</code>. Mayor = más capacidad de largo alcance.",
+    "param.T_train":       "T_train",
+    "param.T_train.tip":   "<strong>Contexto máximo de entrenamiento</strong>. De <code>max_position_embeddings</code>. Más allá es extrapolación.",
+    "param.T_eval":        "T_eval (tu objetivo)",
+    "param.T_eval.tip":    "<strong>Tu contexto de inferencia objetivo</strong>. La pregunta clave: ¿se comportará bien el modelo a ESTA longitud?",
+    "param.n_attn":        "num_attention_heads",
+    "param.n_kv":          "num_key_value_heads",
+    "param.d_head":        "head_dim",
+    "param.n_layers":      "num_hidden_layers",
+    "param.n_params":      "n_params (ej. 8e9)",
+    "param.has_swa":       "¿Tiene SWA?",
+    "common.yes":          "Sí",
+    "common.no":           "No",
+
+    // Tooltips de modos
+    "modes.tip":           "<strong>Cuatro formas de usar la herramienta</strong>.<br><strong>📇 Perfil</strong>: pega un id → las 5 recetas a la vez = TAF Card.<br><strong>🆚 Comparar</strong>: 2-3 modelos lado a lado en una receta.<br><strong>💬 Pregunta</strong>: pregunta libre, el LLM del navegador elige la receta.<br><strong>📋 Receta</strong>: selección manual con control total del formulario.",
+    "profile.tip":         "<strong>Diagnóstico completo en un click</strong>. Pega cualquier id de modelo HF (o elige preset). La herramienta ejecuta las 5 recetas (contexto largo, compresión KV, custom vs API, presupuesto, hardware) y produce una única <strong>TAF Card</strong> con veredicto por dimensión + números clave + clasificación arquitectónica.<br><br><strong>Caso de uso</strong>: \"Estoy evaluando Qwen2.5-32B para producción — ¿cuál es su perfil completo de viabilidad?\" → pega id → Perfilar → listo.",
+    "compare.tip":         "<strong>Misma receta, múltiples modelos</strong>. Elige 2-3 modelos candidatos y una receta. Ve los veredictos en una única tabla comparativa.<br><br><strong>Caso de uso</strong>: \"Necesito recuperación de contexto largo a 16K — ¿cuál es mejor: Llama-3-8B, Mistral-7B o Qwen-7B?\" → elige 3 + X-2 + 16K → ve el ganador.",
+
+    // Modal de ayuda
+    "help.title":               "📘 TAF Agent — Manual de Usuario",
+    "help.what.title":          "¿Qué hace?",
+    "help.what.body":           "Predice la <strong>viabilidad práctica</strong> de cualquier LLM transformer <em>antes de gastar GPU/€</em>. Responde preguntas como \"¿funcionará este modelo a L=32K?\" o \"¿debería entrenar custom o usar API?\" usando fórmulas Python deterministas (TAF — Thermodynamic Attention Framework).",
+    "help.modes.title":         "Cómo usar — 4 modos",
+    "help.modes.profile":       "<strong>📇 Perfilar</strong>: pega id de modelo → todas las recetas a la vez = TAF Card. <strong>Mejor punto de inicio</strong>.",
+    "help.modes.compare":       "<strong>🆚 Comparar</strong>: 2-3 modelos lado a lado en la misma receta. Mejor al elegir entre candidatos.",
+    "help.modes.ask":           "<strong>💬 Pregunta libre</strong>: pregunta en lenguaje natural, el LLM del navegador elige la receta. Mejor para exploración casual.",
+    "help.modes.recipe":        "<strong>📋 Receta + formulario</strong>: selección manual, control total de parámetros. Mejor cuando quieres control exacto.",
+    "help.recipes.title":       "Las 5 recetas disponibles",
+    "help.add_models.title":    "Añadir nuevos modelos (3 maneras)",
+    "help.add_models.preset":   "<strong>Lista de presets</strong>: 11 modelos populares curados. Selecciona del dropdown.",
+    "help.add_models.hf":       "<strong>HF Hub fetch</strong>: pega cualquier id (ej. <code>Qwen/Qwen2.5-32B-Instruct</code>), click 📥 Cargar. El navegador descarga <code>config.json</code> directamente de HuggingFace, llena el formulario. Funciona con cualquier modelo público.",
+    "help.add_models.manual":   "<strong>Manual</strong>: rellena los campos directamente con valores de la model card.",
+    "help.audit.title":         "La cadena auditable",
+    "help.audit.body":          "Cada resultado muestra la <strong>Cadena de Cálculo</strong> completa — cada paso de fórmula con sus entradas, salida e interpretación. Click en cualquier paso para expandir. Las referencias de sección (§26.1, §19.1, etc.) apuntan al paper para la derivación.",
+    "help.synthesis.title":     "La respuesta en lenguaje natural",
+    "help.synthesis.body":      "Tras ejecutar la cadena determinista, un LLM en el navegador (Qwen2.5-0.5B, ~350MB cacheado tras primera carga) sintetiza un resumen en lenguaje natural. Los números arriba son <em>siempre correctos</em> (Python determinista); la síntesis la genera el LLM — verifica contra la cadena si dudas.",
+    "help.params.title":        "Parámetros comunes explicados",
+    "help.verdicts.title":      "Qué mirar en los veredictos",
+    "help.verdict.yes":         "<strong style=\"color:#3fb950;\">SÍ / GO</strong> — procede con confianza; los números apoyan la elección.",
+    "help.verdict.deg":         "<strong style=\"color:#d29922;\">DEGRADADO / TINY-MODEL</strong> — funciona con caveats; lee la acción.",
+    "help.verdict.no":          "<strong style=\"color:#f85149;\">NO / MEMORY-LIMITED</strong> — no procedas tal cual; se da mitigación.",
+    "help.privacy.title":       "Privacidad",
+    "help.privacy.body":        "Todo corre en tu navegador. Sin telemetría, sin analytics, sin datos enviados a ningún sitio. Incluso el modelo LLM corre localmente vía WebGPU/WebAssembly. Tus model_ids y preguntas nunca abandonan esta página.",
+    "help.source.title":        "Código fuente y paper",
+    "help.source.body":         "Código: <a href=\"https://github.com/karlesmarin/tafagent\" target=\"_blank\">github.com/karlesmarin/tafagent</a><br>Paper: <em>Marin 2026 — Transformer Thermodynamics</em> (arXiv próximamente)",
+
     "footer.text":             "© 2026 Carles Marin · Apache-2.0 · investigación independiente · la herramienta que cierra el círculo del paper.",
   },
 
@@ -205,6 +303,55 @@ export const TRANSLATIONS = {
     "profile.hf_placeholder":  "ex. meta-llama/Meta-Llama-3-8B ou Qwen/Qwen2.5-7B",
     "compare.hf_placeholder":  "ID modèle HF (ex. meta-llama/Meta-Llama-3-8B)",
 
+    // Paramètres du formulaire
+    "param.theta":         "θ (rope_theta)",
+    "param.theta.tip":     "<strong>Fréquence de base RoPE</strong> de <code>config.rope_theta</code>. Plus haut = plus de capacité longue portée.",
+    "param.T_train":       "T_train",
+    "param.T_train.tip":   "<strong>Contexte max d'entraînement</strong>. De <code>max_position_embeddings</code>. Au-delà c'est de l'extrapolation.",
+    "param.T_eval":        "T_eval (votre cible)",
+    "param.T_eval.tip":    "<strong>Votre contexte d'inférence cible</strong>. La question clé : le modèle se comportera-t-il bien à CETTE longueur ?",
+    "param.n_attn":        "num_attention_heads",
+    "param.n_kv":          "num_key_value_heads",
+    "param.d_head":        "head_dim",
+    "param.n_layers":      "num_hidden_layers",
+    "param.n_params":      "n_params (ex. 8e9)",
+    "param.has_swa":       "A SWA ?",
+    "common.yes":          "Oui",
+    "common.no":           "Non",
+
+    // Tooltips des modes
+    "modes.tip":           "<strong>Quatre façons d'utiliser l'outil</strong>.<br><strong>📇 Profil</strong>: collez un id → les 5 recettes à la fois = TAF Card.<br><strong>🆚 Comparer</strong>: 2-3 modèles côte à côte sur une recette.<br><strong>💬 Question</strong>: question libre, le LLM du navigateur choisit la recette.<br><strong>📋 Recette</strong>: sélection manuelle avec contrôle total du formulaire.",
+    "profile.tip":         "<strong>Diagnostic complet en un clic</strong>. Collez n'importe quel id de modèle HF (ou choisissez préréglage). L'outil exécute les 5 recettes (contexte long, compression KV, custom vs API, budget, hardware) et produit une <strong>TAF Card</strong> unique avec verdict par dimension + nombres clés + classification architecturale.<br><br><strong>Cas d'usage</strong>: « J'évalue Qwen2.5-32B pour la production — quel est son profil complet de viabilité ? » → collez id → Profiler → fait.",
+    "compare.tip":         "<strong>Même recette, plusieurs modèles</strong>. Choisissez 2-3 modèles candidats et une recette. Voyez les verdicts dans un seul tableau comparatif.<br><br><strong>Cas d'usage</strong>: « J'ai besoin de récupération longue contexte à 16K — quel est le meilleur : Llama-3-8B, Mistral-7B ou Qwen-7B ? » → choisissez 3 + X-2 + 16K → voyez le gagnant.",
+
+    // Modal d'aide
+    "help.title":               "📘 TAF Agent — Manuel d'utilisation",
+    "help.what.title":          "Que fait-il ?",
+    "help.what.body":           "Prédit la <strong>viabilité pratique</strong> de tout LLM transformer <em>avant de dépenser du GPU/€</em>. Répond à des questions comme « ce modèle fonctionnera-t-il à L=32K ? » ou « dois-je entraîner sur mesure ou utiliser une API ? » via des formules Python déterministes (TAF — Thermodynamic Attention Framework).",
+    "help.modes.title":         "Comment l'utiliser — 4 modes",
+    "help.modes.profile":       "<strong>📇 Profiler</strong>: collez id de modèle → toutes les recettes à la fois = TAF Card. <strong>Meilleur point de départ</strong>.",
+    "help.modes.compare":       "<strong>🆚 Comparer</strong>: 2-3 modèles côte à côte sur la même recette. Mieux pour choisir entre candidats.",
+    "help.modes.ask":           "<strong>💬 Question libre</strong>: question en langage naturel, le LLM du navigateur choisit la recette. Mieux pour exploration casuelle.",
+    "help.modes.recipe":        "<strong>📋 Recette + formulaire</strong>: sélection manuelle, contrôle total des paramètres. Mieux quand vous voulez un contrôle exact.",
+    "help.recipes.title":       "Les 5 recettes disponibles",
+    "help.add_models.title":    "Ajouter de nouveaux modèles (3 façons)",
+    "help.add_models.preset":   "<strong>Liste de préréglages</strong>: 11 modèles populaires curés. Sélectionnez dans le dropdown.",
+    "help.add_models.hf":       "<strong>HF Hub fetch</strong>: collez n'importe quel id (ex. <code>Qwen/Qwen2.5-32B-Instruct</code>), cliquez 📥 Charger. Le navigateur télécharge <code>config.json</code> directement de HuggingFace, remplit le formulaire. Fonctionne avec tout modèle public.",
+    "help.add_models.manual":   "<strong>Manuel</strong>: remplissez les champs directement avec les valeurs de la model card.",
+    "help.audit.title":         "La chaîne auditable",
+    "help.audit.body":          "Chaque résultat montre la <strong>Chaîne de Calcul</strong> complète — chaque étape de formule avec ses entrées, sortie et interprétation. Cliquez sur n'importe quelle étape pour développer. Les références de section (§26.1, §19.1, etc.) renvoient au paper pour la dérivation.",
+    "help.synthesis.title":     "La réponse en langage naturel",
+    "help.synthesis.body":      "Après exécution de la chaîne déterministe, un LLM dans le navigateur (Qwen2.5-0.5B, ~350MB cachés après premier chargement) synthétise un résumé en langage naturel. Les nombres ci-dessus sont <em>toujours corrects</em> (Python déterministe) ; la synthèse est générée par LLM — vérifiez contre la chaîne en cas de doute.",
+    "help.params.title":        "Paramètres communs expliqués",
+    "help.verdicts.title":      "Quoi regarder dans les verdicts",
+    "help.verdict.yes":         "<strong style=\"color:#3fb950;\">OUI / GO</strong> — procédez avec confiance ; les nombres soutiennent le choix.",
+    "help.verdict.deg":         "<strong style=\"color:#d29922;\">DÉGRADÉ / TINY-MODEL</strong> — fonctionne avec caveats ; lisez l'action.",
+    "help.verdict.no":          "<strong style=\"color:#f85149;\">NON / MEMORY-LIMITED</strong> — ne procédez pas tel quel ; mitigation fournie.",
+    "help.privacy.title":       "Confidentialité",
+    "help.privacy.body":        "Tout s'exécute dans votre navigateur. Pas de télémétrie, pas d'analytique, pas de données envoyées ailleurs. Même le modèle LLM s'exécute localement via WebGPU/WebAssembly. Vos model_ids et questions ne quittent jamais cette page.",
+    "help.source.title":        "Code source et paper",
+    "help.source.body":         "Code : <a href=\"https://github.com/karlesmarin/tafagent\" target=\"_blank\">github.com/karlesmarin/tafagent</a><br>Paper : <em>Marin 2026 — Transformer Thermodynamics</em> (arXiv à venir)",
+
     "footer.text":             "© 2026 Carles Marin · Apache-2.0 · recherche indépendante · l'outil qui ferme la boucle du paper.",
   },
 
@@ -271,6 +418,55 @@ export const TRANSLATIONS = {
     "profile.hf_placeholder":  "例如: meta-llama/Meta-Llama-3-8B 或 Qwen/Qwen2.5-7B",
     "compare.hf_placeholder":  "HF 模型 id (例如: meta-llama/Meta-Llama-3-8B)",
 
+    // 表单参数
+    "param.theta":         "θ (rope_theta)",
+    "param.theta.tip":     "<strong>RoPE 基础频率</strong> 来自 <code>config.rope_theta</code>。越高 = 长程能力越强。",
+    "param.T_train":       "T_train",
+    "param.T_train.tip":   "<strong>训练最大上下文</strong>。来自 <code>max_position_embeddings</code>。超出此范围属于外推。",
+    "param.T_eval":        "T_eval (您的目标)",
+    "param.T_eval.tip":    "<strong>您的目标推理上下文</strong>。关键问题: 模型在 <em>这个</em> 长度下表现是否良好?",
+    "param.n_attn":        "num_attention_heads",
+    "param.n_kv":          "num_key_value_heads",
+    "param.d_head":        "head_dim",
+    "param.n_layers":      "num_hidden_layers",
+    "param.n_params":      "n_params (例如 8e9)",
+    "param.has_swa":       "有 SWA 吗?",
+    "common.yes":          "是",
+    "common.no":           "否",
+
+    // 模式提示
+    "modes.tip":           "<strong>四种使用方式</strong>。<br><strong>📇 画像</strong>: 粘贴模型 id → 一次运行所有 5 个配方 = TAF 卡。<br><strong>🆚 比较</strong>: 2-3 个模型在一个配方上并排比较。<br><strong>💬 提问</strong>: 自由形式问题,浏览器 LLM 选择配方。<br><strong>📋 配方</strong>: 手动选择,完全控制表单。",
+    "profile.tip":         "<strong>一键完整诊断</strong>。粘贴任意 HF 模型 id (或选择预设)。工具运行所有 5 个配方 (长上下文、KV 压缩、自定义 vs API、预算、硬件),生成单个 <strong>TAF 卡</strong>,显示每个维度的判定 + 关键数字 + 架构分类。<br><br><strong>用例</strong>: \"我正在为生产评估 Qwen2.5-32B — 它的完整可行性概况是什么?\" → 粘贴 id → 画像 → 完成。",
+    "compare.tip":         "<strong>同一配方,多个模型</strong>。选择 2-3 个候选模型和一个配方。在单个比较表中查看判定。<br><br><strong>用例</strong>: \"我需要在 16K 进行长上下文检索 — 哪个最好: Llama-3-8B、Mistral-7B 或 Qwen-7B?\" → 选择 3 个 + X-2 + 16K → 看赢家。",
+
+    // 帮助模态框
+    "help.title":               "📘 TAF Agent — 用户手册",
+    "help.what.title":          "它做什么?",
+    "help.what.body":           "在<em>花费 GPU/$ 之前</em>,预测任意 transformer LLM 的<strong>实际可行性</strong>。回答诸如 \"这个模型能在 L=32K 工作吗?\" 或 \"我应该自定义训练还是使用 API?\" 等问题,使用确定性 Python 公式 (TAF — Thermodynamic Attention Framework)。",
+    "help.modes.title":         "如何使用 — 4 种模式",
+    "help.modes.profile":       "<strong>📇 画像</strong>: 粘贴模型 id → 同时运行所有配方 = TAF 卡。<strong>最佳起点</strong>。",
+    "help.modes.compare":       "<strong>🆚 比较</strong>: 2-3 个模型在同一配方上并排。最适合在候选者之间选择。",
+    "help.modes.ask":           "<strong>💬 自由提问</strong>: 自然语言问题,浏览器 LLM 选择配方。最适合随意探索。",
+    "help.modes.recipe":        "<strong>📋 配方 + 表单</strong>: 手动选择,完全控制参数。最适合需要精确控制时。",
+    "help.recipes.title":       "可用的 5 个配方",
+    "help.add_models.title":    "添加新模型 (3 种方式)",
+    "help.add_models.preset":   "<strong>预设列表</strong>: 11 个流行模型已策划。从下拉菜单选择。",
+    "help.add_models.hf":       "<strong>HF Hub 获取</strong>: 粘贴任意 id (例如 <code>Qwen/Qwen2.5-32B-Instruct</code>),点击 📥 获取。浏览器直接从 HuggingFace 下载 <code>config.json</code>,填充表单。适用于任何公共模型。",
+    "help.add_models.manual":   "<strong>手动</strong>: 用模型卡的值直接填充表单字段。",
+    "help.audit.title":         "可审计链",
+    "help.audit.body":          "每个结果都显示完整的<strong>计算链</strong> — 每个公式步骤及其输入、输出和解释。点击任意步骤展开。引用的章节号 (§26.1、§19.1 等) 指向论文中的推导。",
+    "help.synthesis.title":     "自然语言回答",
+    "help.synthesis.body":      "在确定性链运行后,浏览器中的 LLM (Qwen2.5-0.5B,首次加载后约 350MB 缓存) 综合自然语言摘要。上面的数字<em>始终正确</em> (确定性 Python);综合由 LLM 生成 — 如有疑问,请对照链验证。",
+    "help.params.title":        "常见参数解释",
+    "help.verdicts.title":      "判定中要看什么",
+    "help.verdict.yes":         "<strong style=\"color:#3fb950;\">是 / GO</strong> — 自信地继续;数字支持选择。",
+    "help.verdict.deg":         "<strong style=\"color:#d29922;\">降级 / TINY-MODEL</strong> — 有警告地工作;阅读操作。",
+    "help.verdict.no":          "<strong style=\"color:#f85149;\">否 / MEMORY-LIMITED</strong> — 不要按原样进行;提供缓解措施。",
+    "help.privacy.title":       "隐私",
+    "help.privacy.body":        "一切都在您的浏览器中运行。无遥测,无分析,无数据发送到任何地方。即使是 LLM 模型也通过 WebGPU/WebAssembly 在本地运行。您的 model_ids 和问题永不离开此页面。",
+    "help.source.title":        "源代码和论文",
+    "help.source.body":         "源代码: <a href=\"https://github.com/karlesmarin/tafagent\" target=\"_blank\">github.com/karlesmarin/tafagent</a><br>论文: <em>Marin 2026 — Transformer Thermodynamics</em> (arXiv 即将)",
+
     "footer.text":             "© 2026 Carles Marin · Apache-2.0 · 独立研究 · 闭合论文回路的工具。",
   },
 };