v0.7.0: SWA Unmasker (anti-bullshit #1) + foldable main panels + preset auto-fill

Ships the first feature of the v0.7 anti-bullshit pack inspired by HF community pain points: detect when a model card claims a max_position_embeddings far larger than its effective context (Mistral-7B-v0.1: declared 32k, attends ~8k via SWA).

NEW
- 🪟 Unmask mode: paste an HF model id (or raw config.json) → 1-second verdict (HONEST / INFLATED / SEVERELY INFLATED / YARN-EXTENDED). Pure browser arithmetic on config.json: SWA window + RoPE-scaling + GQA/d_head heuristics. No GPU, no inference.
- js/swa_unmasker.js: pure logic module, no human strings. Returns warning codes + params; main.js renders via i18n with {placeholder} substitution so EN/ES/FR/ZH all work.
- tFmt() i18n helper: t(key) with {placeholder} replacement.

UX
- All <main> sections now wrapped at runtime in <details open> with foldable header (idempotent wrapMainSectionsAsFoldable). Big ▼ arrow rotates to ▶ when collapsed. Triple-browser marker hide (list-style + ::-webkit-details-marker + ::marker).
- Inventory modal: 4 inv-cards now <details open> with arrow per card.
- Architectures-supported defaults to open; tooltip text removed "(click to expand)".
- Profile + Recipe preset dropdowns now also auto-fill the HF id input (presets are HF model ids; clarifies dual source of truth — preset = cached, 📥 Fetch = live HF Hub).

i18n cleanup (the big one)
- swa_unmasker hardcoded English strings: GONE. All warnings/recommendations/verdict labels live in i18n with {placeholder} substitution × EN/ES/FR/ZH.
- 8 mode-desc strings: refactored from inline string switch to t(`mode_desc.${mode}`) lookup. Section show/hide reduced to a sectionMap object.
- modes.tip updated: "7 modes" → "8 modes" (added Unmask).
- 423 keys × 4 langs, 0 missing / 0 extra (parity verified).

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

index.html

@@ -249,8 +249,8 @@
       <button class="help-close" id="inventory-close" aria-label="Close inventory">×</button>
       <h2 id="inv-modal-title" data-i18n="inv.title">🧰 What this tool gives you</h2>
       <div class="inventory-grid">
-        <div class="inv-card">
-          <h3 data-i18n="inv.recipes.title">🎯 8 recipes &mdash; does this model fit your use case?</h3>
+        <details class="inv-card" open>
+          <summary class="inv-card-title" data-i18n="inv.recipes.title">🎯 8 recipes &mdash; does this model fit your use case?</summary>
           <ul>
             <li><strong data-i18n="inv.recipes.x1.title">Custom train vs API</strong>: <span data-i18n="inv.recipes.x1.body">which is cheaper for your traffic?</span></li>
             <li><strong data-i18n="inv.recipes.x2.title">Long context</strong>: <span data-i18n="inv.recipes.x2.body">will it handle 32k / 128k tokens reliably?</span></li>
@@ -261,35 +261,35 @@
             <li><strong data-i18n="inv.recipes.x22.title">Compute-context</strong>: <span data-i18n="inv.recipes.x22.body">does the model fit the empirical band?</span></li>
             <li><strong data-i18n="inv.recipes.x23.title">IH-phase</strong>: <span data-i18n="inv.recipes.x23.body">pre- or post-induction-head?</span></li>
           </ul>
-        </div>
-        <div class="inv-card">
-          <h3 data-i18n="inv.diag.title">🔬 Diagnostics</h3>
+        </details>
+        <details class="inv-card" open>
+          <summary class="inv-card-title" data-i18n="inv.diag.title">🔬 Diagnostics</summary>
           <ul>
             <li data-i18n="inv.diag.gamma"><strong>γ predicted vs observed</strong> &mdash; auto-classifies the model into 5 regimes (normal · fraud / inflated context · compressed · over-Padé · sliding-window)</li>
             <li data-i18n="inv.diag.cardy"><strong>Cardy ΔH</strong> &mdash; entropy shift between observed and nominal context</li>
             <li data-i18n="inv.diag.fals"><strong>Falsification dashboard</strong> &mdash; checks 23 specific predictions (F1–F23)</li>
             <li data-i18n="inv.diag.alg"><strong>Algebraic consistency</strong> &mdash; 8 mathematical identities the model must satisfy</li>
           </ul>
-        </div>
-        <div class="inv-card">
-          <h3 data-i18n="inv.verify.title">✓ Formally verified math</h3>
+        </details>
+        <details class="inv-card" open>
+          <summary class="inv-card-title" data-i18n="inv.verify.title">✓ Formally verified math</summary>
           <ul>
             <li data-i18n="inv.verify.count"><strong>37 theorems</strong> machine-proven in Lean 4 + Mathlib4</li>
             <li data-i18n="inv.verify.click">Click any badge → opens the source line on GitHub</li>
             <li data-i18n="inv.verify.reverify">Verify yourself: <code>lake build</code> (≈5 s after cache fetch)</li>
           </ul>
-        </div>
-        <div class="inv-card">
-          <h3 data-i18n="inv.export.title">📤 Export &amp; share</h3>
+        </details>
+        <details class="inv-card" open>
+          <summary class="inv-card-title" data-i18n="inv.export.title">📤 Export &amp; share</summary>
           <ul>
             <li data-i18n="inv.export.formats"><strong>JSON · Markdown · LaTeX</strong> (paper-ready)</li>
             <li data-i18n="inv.export.share">Reproducible share link (state encoded in URL)</li>
             <li data-i18n="inv.export.registry">Submit to community registry on GitHub</li>
           </ul>
-        </div>
+        </details>
       </div>
 
-      <details class="arch-supported">
+      <details class="arch-supported" open>
         <summary data-i18n="arch.summary">Architectures supported (click to expand)</summary>
         <div class="arch-badges">
           <span class="badge">✓ RoPE-MHA <span class="info"><span class="tooltip" data-i18n="tooltip.mha">Multi-Head Attention: each token position attends through several parallel heads at once.</span></span></span>
@@ -334,6 +334,7 @@
         <button class="mode-btn" data-mode="recipe" role="tab" aria-selected="false" data-i18n="modes.recipe">📋 Pick recipe</button>
         <button class="mode-btn" data-mode="diagnose" role="tab" aria-selected="false" data-i18n="modes.diagnose">🩺 Diagnose CLI</button>
         <button class="mode-btn" data-mode="phase" role="tab" aria-selected="false" data-i18n="modes.phase">📊 Phase diagram</button>
+        <button class="mode-btn" data-mode="unmask" role="tab" aria-selected="false" data-i18n="modes.unmask">🪟 Unmask</button>
       </div>
       <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>),
@@ -623,6 +624,34 @@
       <div id="phase-info" class="recipe-desc" style="margin-top:0.6em;"></div>
     </section>
 
+    <!-- Unmask mode: detect misleading max_position_embeddings via SWA / RoPE-scaling -->
+    <section id="unmask-section" style="display:none;">
+      <h2><span data-i18n="unmask.title">🪟 Context Unmasker</span>
+        <span class="info"><span class="tooltip" data-i18n="unmask.tip">
+          Paste a HuggingFace model id (or raw config.json). The tool checks for
+          sliding-window attention, RoPE scaling (YaRN/linear/dynamic NTK), and
+          GQA — anything that makes <code>max_position_embeddings</code> larger
+          than the practical effective context. Mistral-7B-v0.1 is the canonical
+          example: declared 32k, attends within ~4-8k.
+        </span></span>
+      </h2>
+      <p class="recipe-desc" data-i18n="unmask.desc">
+        <strong>Are you about to spend money on a model that won't actually attend that far?</strong> Paste an id and find out in 1 second. No GPU, no inference — just config.json arithmetic.
+      </p>
+      <div class="form-row">
+        <label for="unmask-id" data-i18n="unmask.id_label">HF model id:</label>
+        <input type="text" id="unmask-id" placeholder="e.g. mistralai/Mistral-7B-v0.1" />
+        <button type="button" id="unmask-fetch-btn" data-i18n="unmask.fetch_btn">🔍 Unmask</button>
+      </div>
+      <p id="unmask-status" class="recipe-desc" style="font-size:0.92em;"></p>
+      <details style="margin: 0.6em 0;">
+        <summary style="cursor:pointer; font-size:0.92em;" data-i18n="unmask.paste_summary">Or paste raw config.json (private / in-dev models)</summary>
+        <textarea id="unmask-paste" rows="6" style="width:100%; font-family:monospace; font-size:0.85em; margin-top:0.4em;" placeholder='{"max_position_embeddings": 32768, "sliding_window": 4096, ...}'></textarea>
+        <button type="button" id="unmask-paste-btn" data-i18n="unmask.paste_btn" style="margin-top:0.4em;">🔍 Unmask pasted config</button>
+      </details>
+      <div id="unmask-output" style="margin-top: 1em;"></div>
+    </section>
+
     <!-- Recipe selector (mode=recipe) -->
     <section id="recipe-section" style="display:none;">
       <h2 data-i18n="recipe.title">📋 Recipe</h2>

js/i18n.js

@@ -125,7 +125,7 @@ export const TRANSLATIONS = {
     "inv.export.formats":          "<strong>JSON · Markdown · LaTeX</strong> (paper-ready)",
     "inv.export.share":            "Reproducible share link (state encoded in URL)",
     "inv.export.registry":         "Submit to community registry on GitHub",
-    "arch.summary":                "Architectures supported (click to expand)",
+    "arch.summary":                "Architectures supported",
     "arch.anyhf":                  "✓ Any HuggingFace public model",
     "tooltip.mha":                 "Multi-Head Attention: each token position attends through several parallel heads at once.",
     "tooltip.gqa":                 "Grouped Query Attention: queries share fewer keys/values than heads (saves memory but pushes γ toward Hagedorn).",
@@ -133,6 +133,62 @@ export const TRANSLATIONS = {
     "tooltip.abspe":               "Absolute Position Embeddings: each position has a fixed learned vector added to the token embedding.",
     "tooltip.swa":                 "Sliding Window Attention: each token only attends within a fixed local window (Mistral, gemma-2 use this).",
     "tooltip.ssm":                 "State Space Model: a sequence layer that maintains internal state instead of attention (Mamba, Jamba use this).",
+
+    // v0.7.0 — anti-bullshit pack #1: SWA / RoPE-scaling unmasker
+    "modes.unmask":                "🪟 Unmask",
+    "unmask.title":                "🪟 Context Unmasker",
+    "unmask.tip":                  "Paste a HuggingFace model id (or raw config.json). The tool checks for sliding-window attention, RoPE scaling (YaRN/linear/dynamic NTK), and GQA — anything that makes <code>max_position_embeddings</code> larger than the practical effective context. Mistral-7B-v0.1 is the canonical example: declared 32k, attends within ~4-8k.",
+    "unmask.desc":                 "<strong>Are you about to spend money on a model that won't actually attend that far?</strong> Paste an id and find out in 1 second. No GPU, no inference — just config.json arithmetic.",
+    "unmask.id_label":             "HF model id:",
+    "unmask.fetch_btn":            "🔍 Unmask",
+    "unmask.paste_summary":        "Or paste raw config.json (private / in-dev models)",
+    "unmask.paste_btn":            "🔍 Unmask pasted config",
+    "unmask.label.declared":       "Declared context",
+    "unmask.label.effective":      "Effective (estimate)",
+    "unmask.label.ratio":          "Ratio",
+    "unmask.section.flags":        "Architecture flags",
+    "unmask.section.warnings":     "Warnings",
+    "unmask.section.reco":         "Recommendation",
+    "unmask.flag.swa":             "SWA",
+    "unmask.flag.rope":            "RoPE scaling",
+    "unmask.flag.gqa":             "GQA",
+    "unmask.flag.layers":          "Layers",
+    "unmask.flag.dhead":           "d_head",
+    "unmask.flag.theta":           "RoPE θ",
+    "unmask.flag.yes":             "yes",
+    "unmask.flag.no":              "no",
+    "unmask.flag.full_mha":        "no (full MHA, {n} heads)",
+    "unmask.verdict.honest":            "✅ HONEST",
+    "unmask.verdict.inflated":          "⚠ INFLATED",
+    "unmask.verdict.severely_inflated": "❌ SEVERELY INFLATED",
+    "unmask.verdict.yarn_extended":     "⚠ YARN-EXTENDED",
+    "unmask.verdict.unknown":           "❓ UNKNOWN",
+    "unmask.warn.swa_window":      "SWA window: {window} tokens — each layer only attends within this window.",
+    "unmask.warn.multihop":        "Multi-hop estimate: ~{multiHop} tokens (conservative: window × {factor}).",
+    "unmask.warn.yarn":            "RoPE scaling ({type}) extends context {factor}× from ~{original} to {declared} tokens.",
+    "unmask.warn.yarn_advice":     "RoPE-extended context — verify γ behavior at the full claimed length with the γ_check diagnostic.",
+    "unmask.warn.gqa_small_dhead": "Small head dim ({d_head}) + GQA: KV cache compression at long context is likely (γ pushed toward Hagedorn).",
+    "unmask.reco.honest":              "Standard full-attention model. Effective context matches declared ({declared} tokens).",
+    "unmask.reco.inflated":            "Effective ~{effective} tokens via SWA. Use γ_check to verify behavior at your target evaluation length.",
+    "unmask.reco.severely_inflated":   "Treat as a ~{effective}-token context model in practice. The {declared}-token claim only applies via cross-layer attention chains, which empirically degrade past ~2× the SWA window.",
+    "unmask.reco.yarn_extended":       "RoPE-extended context. Run a long-context benchmark (NIAH at 8k / 16k / 32k / full) to confirm the extension holds. Use γ_check with T_eval = {declared}.",
+    "unmask.reco.unknown":             "Could not parse config. Verify the URL is a valid HF model with public config.json.",
+    "unmask.status.empty_id":      "⚠ Enter a model id (e.g. mistralai/Mistral-7B-v0.1).",
+    "unmask.status.fetching":      "⏳ Fetching config.json for {modelId}...",
+    "unmask.status.success":       "✅ Analyzed {modelId} (verdict: {verdict})",
+    "unmask.status.empty_paste":   "⚠ Paste a config.json first.",
+    "unmask.status.invalid_json":  "❌ Not valid JSON: {error}",
+    "unmask.status.success_paste": "✅ Analyzed pasted config (verdict: {verdict})",
+    "unmask.pasted_label":         "(pasted config)",
+    "mode_desc.ask":               "Type a free-form question. The in-browser LLM picks the right recipe and runs it.",
+    "mode_desc.recipe":            "Pick a recipe directly and fill the form. Full manual control.",
+    "mode_desc.profile":           "Quickest start: paste any HuggingFace model id, click Profile. See all 5 recipes scored in seconds.",
+    "mode_desc.compare":           "Pick 2-3 candidate models + one recipe. See verdicts side-by-side in a comparison table.",
+    "mode_desc.inspector":         "Paste a config.json directly. Useful for private/in-development models not on HF Hub.",
+    "mode_desc.diagnose":          "Build the diagnose_model.py CLI command to MEASURE γ_obs on real GPU. Browser predicts; CLI measures.",
+    "mode_desc.phase":             "γ × θ scatter of the paper's empirical panel. Hover a dot for details, click to load into Diagnose / Recipe forms.",
+    "mode_desc.unmask":            "Detects whether max_position_embeddings is misleading (SWA / YaRN / RoPE-scaling). Paste a model id, get a 1-line verdict.",
+    "profile.preset_loaded":       "✅ Loaded preset for <strong>{id}</strong>. Form pre-filled. (Click 📥 Fetch to override with the latest config from HF Hub.)",
     "share.import_desc":       "Got a JSON file from someone else's TAF analysis? Load it here to see the verdict + chain locally. Same view as if you'd run it yourself.",
     "share.import_btn":        "📂 Load shared JSON",
     "synthesis.system":        "You are a precise transformer LLM diagnostic assistant. Given pre-computed TAF formula results, write a clear plain-English summary in 4-6 sentences. Cite the section number (§X.Y) for each number you mention. Always give a concrete recommendation. Do NOT invent numbers.",
@@ -225,7 +281,7 @@ export const TRANSLATIONS = {
     "common.no":           "No",
 
     // Mode tooltips
-    "modes.tip":           "<strong>Seven ways to use the tool</strong>.<br><strong>📇 Profile</strong>: paste a model id → 5-recipe TAF Card.<br><strong>🆚 Compare</strong>: 2-3 models side-by-side on one recipe.<br><strong>🔍 Inspect config</strong>: paste raw config.json → full Profile.<br><strong>💬 Ask</strong>: free-form question, browser LLM picks the recipe.<br><strong>📋 Recipe</strong>: manual selection with full form control.<br><strong>🩺 Diagnose CLI</strong>: generate Python command for local γ measurement.<br><strong>📊 Phase diagram</strong>: 23-model panel on (log θ, γ) plane.",
+    "modes.tip":           "<strong>Eight ways to use the tool</strong>.<br><strong>📇 Profile</strong>: paste a model id → 5-recipe TAF Card.<br><strong>🆚 Compare</strong>: 2-3 models side-by-side on one recipe.<br><strong>🔍 Inspect config</strong>: paste raw config.json → full Profile.<br><strong>💬 Ask</strong>: free-form question, browser LLM picks the recipe.<br><strong>📋 Recipe</strong>: manual selection with full form control.<br><strong>🩺 Diagnose CLI</strong>: generate Python command for local γ measurement.<br><strong>📊 Phase diagram</strong>: 23-model panel on (log θ, γ) plane.<br><strong>🪟 Unmask</strong>: detect misleading max_position_embeddings (SWA / YaRN / RoPE-scaling).",
     "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.",
 
@@ -682,7 +738,7 @@ export const TRANSLATIONS = {
     "inv.export.formats":          "<strong>JSON · Markdown · LaTeX</strong> (listo para paper)",
     "inv.export.share":            "Link reproducible (estado codificado en URL)",
     "inv.export.registry":         "Envía al registro comunitario en GitHub",
-    "arch.summary":                "Arquitecturas soportadas (click para expandir)",
+    "arch.summary":                "Arquitecturas soportadas",
     "arch.anyhf":                  "✓ Cualquier modelo público de HuggingFace",
     "tooltip.mha":                 "Multi-Head Attention: cada posición atiende mediante varios heads paralelos a la vez.",
     "tooltip.gqa":                 "Grouped Query Attention: las queries comparten menos keys/values que heads (ahorra memoria pero empuja γ hacia Hagedorn).",
@@ -690,6 +746,62 @@ export const TRANSLATIONS = {
     "tooltip.abspe":               "Absolute Position Embeddings: cada posición tiene un vector fijo aprendido sumado al embedding del token.",
     "tooltip.swa":                 "Sliding Window Attention: cada token solo atiende dentro de una ventana local fija (Mistral, gemma-2 lo usan).",
     "tooltip.ssm":                 "State Space Model: capa de secuencia que mantiene estado interno en lugar de atención (Mamba, Jamba lo usan).",
+
+    // v0.7.0 — anti-bullshit pack #1: SWA / RoPE-scaling unmasker
+    "modes.unmask":                "🪟 Desenmascarar",
+    "unmask.title":                "🪟 Desenmascarador de contexto",
+    "unmask.tip":                  "Pega un id de modelo HuggingFace (o config.json crudo). La herramienta detecta sliding-window attention, RoPE scaling (YaRN/linear/dynamic NTK), y GQA — todo lo que hace que <code>max_position_embeddings</code> sea mayor que el contexto efectivo real. Mistral-7B-v0.1 es el ejemplo canónico: declara 32k, atiende dentro de ~4-8k.",
+    "unmask.desc":                 "<strong>¿Estás a punto de gastar dinero en un modelo que en realidad no atiende tan lejos?</strong> Pega un id y descúbrelo en 1 segundo. Sin GPU, sin inferencia — solo aritmética sobre config.json.",
+    "unmask.id_label":             "ID modelo HF:",
+    "unmask.fetch_btn":            "🔍 Desenmascarar",
+    "unmask.paste_summary":        "O pega config.json crudo (modelos privados / en desarrollo)",
+    "unmask.paste_btn":            "🔍 Desenmascarar config pegado",
+    "unmask.label.declared":       "Contexto declarado",
+    "unmask.label.effective":      "Efectivo (estimado)",
+    "unmask.label.ratio":          "Ratio",
+    "unmask.section.flags":        "Banderas de arquitectura",
+    "unmask.section.warnings":     "Avisos",
+    "unmask.section.reco":         "Recomendación",
+    "unmask.flag.swa":             "SWA",
+    "unmask.flag.rope":            "RoPE scaling",
+    "unmask.flag.gqa":             "GQA",
+    "unmask.flag.layers":          "Capas",
+    "unmask.flag.dhead":           "d_head",
+    "unmask.flag.theta":           "RoPE θ",
+    "unmask.flag.yes":             "sí",
+    "unmask.flag.no":              "no",
+    "unmask.flag.full_mha":        "no (MHA completo, {n} heads)",
+    "unmask.verdict.honest":            "✅ HONESTO",
+    "unmask.verdict.inflated":          "⚠ INFLADO",
+    "unmask.verdict.severely_inflated": "❌ GRAVEMENTE INFLADO",
+    "unmask.verdict.yarn_extended":     "⚠ YARN-EXTENDIDO",
+    "unmask.verdict.unknown":           "❓ DESCONOCIDO",
+    "unmask.warn.swa_window":      "Ventana SWA: {window} tokens — cada capa solo atiende dentro de esta ventana.",
+    "unmask.warn.multihop":        "Estimación multi-hop: ~{multiHop} tokens (conservador: ventana × {factor}).",
+    "unmask.warn.yarn":            "RoPE scaling ({type}) extiende contexto {factor}× desde ~{original} hasta {declared} tokens.",
+    "unmask.warn.yarn_advice":     "Contexto RoPE-extendido — verifica el comportamiento de γ a la longitud declarada con el diagnóstico γ_check.",
+    "unmask.warn.gqa_small_dhead": "head dim pequeño ({d_head}) + GQA: probable compresión de KV cache a contexto largo (γ empujado hacia Hagedorn).",
+    "unmask.reco.honest":              "Modelo de atención completa estándar. Contexto efectivo coincide con declarado ({declared} tokens).",
+    "unmask.reco.inflated":            "Efectivo ~{effective} tokens vía SWA. Usa γ_check para verificar el comportamiento a tu longitud objetivo.",
+    "unmask.reco.severely_inflated":   "Trátalo como un modelo de ~{effective} tokens en la práctica. El claim de {declared} tokens solo aplica vía cadenas de atención cross-layer, que empíricamente degradan más allá de ~2× la ventana SWA.",
+    "unmask.reco.yarn_extended":       "Contexto RoPE-extendido. Corre un benchmark long-context (NIAH a 8k / 16k / 32k / full) para confirmar que la extensión se sostiene. Usa γ_check con T_eval = {declared}.",
+    "unmask.reco.unknown":             "No se pudo parsear el config. Verifica que la URL sea un modelo HF válido con config.json público.",
+    "unmask.status.empty_id":      "⚠ Introduce un model id (ej. mistralai/Mistral-7B-v0.1).",
+    "unmask.status.fetching":      "⏳ Obteniendo config.json para {modelId}...",
+    "unmask.status.success":       "✅ Analizado {modelId} (veredicto: {verdict})",
+    "unmask.status.empty_paste":   "⚠ Pega un config.json primero.",
+    "unmask.status.invalid_json":  "❌ JSON inválido: {error}",
+    "unmask.status.success_paste": "✅ Config pegado analizado (veredicto: {verdict})",
+    "unmask.pasted_label":         "(config pegado)",
+    "mode_desc.ask":               "Escribe una pregunta libre. El LLM en el navegador elige la receta correcta y la ejecuta.",
+    "mode_desc.recipe":            "Selecciona una receta directamente y rellena el formulario. Control manual completo.",
+    "mode_desc.profile":           "Inicio más rápido: pega cualquier model id de HuggingFace, click Profile. Mira las 5 recetas en segundos.",
+    "mode_desc.compare":           "Elige 2-3 modelos candidatos + una receta. Ve veredictos lado a lado en tabla.",
+    "mode_desc.inspector":         "Pega un config.json directamente. Útil para modelos privados / en desarrollo no en HF Hub.",
+    "mode_desc.diagnose":          "Construye el comando CLI diagnose_model.py para MEDIR γ_obs en GPU real. El navegador predice; el CLI mide.",
+    "mode_desc.phase":             "Scatter γ × θ del panel empírico del paper. Hover sobre puntos para detalles, click para cargar en Diagnose / Recipe.",
+    "mode_desc.unmask":            "Detecta si max_position_embeddings es engañoso (SWA / YaRN / RoPE-scaling). Pega un model id, obtén un veredicto en 1 línea.",
+    "profile.preset_loaded":       "✅ Preset cargado para <strong>{id}</strong>. Formulario pre-rellenado. (Click 📥 Fetch para sobreescribir con el último config de HF Hub.)",
     "share.import_desc":       "¿Tienes un fichero JSON del análisis TAF de alguien? Cárgalo aquí para ver el veredicto + cadena localmente. La misma vista que si lo hubieras ejecutado tú.",
     "share.import_btn":        "📂 Cargar JSON compartido",
     "synthesis.system":        "Eres un asistente de diagnóstico preciso para LLMs transformer. Dados resultados de fórmulas TAF pre-calculados, escribe un resumen claro en español de 4-6 frases. Cita el número de sección (§X.Y) para cada número que menciones. Da siempre una recomendación concreta. NO inventes números.",
@@ -782,7 +894,7 @@ export const TRANSLATIONS = {
     "common.no":           "No",
 
     // Tooltips de modos
-    "modes.tip":           "<strong>Siete formas de usar la herramienta</strong>.<br><strong>📇 Perfil</strong>: pega un id → TAF Card de 5 recetas.<br><strong>🆚 Comparar</strong>: 2-3 modelos lado a lado en una receta.<br><strong>🔍 Inspeccionar config</strong>: pega config.json crudo → Perfil completo.<br><strong>💬 Pregunta</strong>: pregunta libre, el LLM del navegador elige la receta.<br><strong>📋 Receta</strong>: selección manual con control total del formulario.<br><strong>🩺 Diagnóstico CLI</strong>: genera comando Python para medir γ localmente.<br><strong>📊 Diagrama de fase</strong>: panel de 23 modelos en plano (log θ, γ).",
+    "modes.tip":           "<strong>Ocho formas de usar la herramienta</strong>.<br><strong>📇 Perfil</strong>: pega un id → TAF Card de 5 recetas.<br><strong>🆚 Comparar</strong>: 2-3 modelos lado a lado en una receta.<br><strong>🔍 Inspeccionar config</strong>: pega config.json crudo → Perfil completo.<br><strong>💬 Pregunta</strong>: pregunta libre, el LLM del navegador elige la receta.<br><strong>📋 Receta</strong>: selección manual con control total del formulario.<br><strong>🩺 Diagnóstico CLI</strong>: genera comando Python para medir γ localmente.<br><strong>📊 Diagrama de fase</strong>: panel de 23 modelos en plano (log θ, γ).<br><strong>🪟 Desenmascarar</strong>: detecta max_position_embeddings engañoso (SWA / YaRN / RoPE-scaling).",
     "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.",
 
@@ -1103,7 +1215,7 @@ export const TRANSLATIONS = {
     "inv.export.formats":          "<strong>JSON · Markdown · LaTeX</strong> (prêt pour papier)",
     "inv.export.share":            "Lien reproductible (état encodé dans l'URL)",
     "inv.export.registry":         "Soumettre au registre communautaire sur GitHub",
-    "arch.summary":                "Architectures prises en charge (cliquez pour déplier)",
+    "arch.summary":                "Architectures prises en charge",
     "arch.anyhf":                  "✓ Tout modèle public HuggingFace",
     "tooltip.mha":                 "Multi-Head Attention : chaque position attend via plusieurs têtes parallèles à la fois.",
     "tooltip.gqa":                 "Grouped Query Attention : les queries partagent moins de keys/values que de heads (économise mémoire mais pousse γ vers Hagedorn).",
@@ -1111,6 +1223,62 @@ export const TRANSLATIONS = {
     "tooltip.abspe":               "Absolute Position Embeddings : chaque position a un vecteur fixe appris ajouté au token.",
     "tooltip.swa":                 "Sliding Window Attention : chaque token n'attend que dans une fenêtre locale fixe (Mistral, gemma-2 l'utilisent).",
     "tooltip.ssm":                 "State Space Model : couche de séquence qui maintient un état interne au lieu d'attention (Mamba, Jamba l'utilisent).",
+
+    // v0.7.0 — anti-bullshit pack #1: SWA / RoPE-scaling unmasker
+    "modes.unmask":                "🪟 Démasquer",
+    "unmask.title":                "🪟 Démasqueur de contexte",
+    "unmask.tip":                  "Collez un id de modèle HuggingFace (ou config.json brut). L'outil détecte sliding-window attention, RoPE scaling (YaRN/linear/dynamic NTK), et GQA — tout ce qui rend <code>max_position_embeddings</code> plus grand que le contexte effectif réel. Mistral-7B-v0.1 est l'exemple canonique : déclare 32k, attend dans ~4-8k.",
+    "unmask.desc":                 "<strong>Êtes-vous sur le point de dépenser de l'argent sur un modèle qui n'attend pas vraiment aussi loin ?</strong> Collez un id et découvrez-le en 1 seconde. Sans GPU, sans inférence — juste de l'arithmétique sur config.json.",
+    "unmask.id_label":             "ID modèle HF :",
+    "unmask.fetch_btn":            "🔍 Démasquer",
+    "unmask.paste_summary":        "Ou collez config.json brut (modèles privés / en dev)",
+    "unmask.paste_btn":            "🔍 Démasquer config collé",
+    "unmask.label.declared":       "Contexte déclaré",
+    "unmask.label.effective":      "Effectif (estimé)",
+    "unmask.label.ratio":          "Ratio",
+    "unmask.section.flags":        "Drapeaux d'architecture",
+    "unmask.section.warnings":     "Avertissements",
+    "unmask.section.reco":         "Recommandation",
+    "unmask.flag.swa":             "SWA",
+    "unmask.flag.rope":            "RoPE scaling",
+    "unmask.flag.gqa":             "GQA",
+    "unmask.flag.layers":          "Couches",
+    "unmask.flag.dhead":           "d_head",
+    "unmask.flag.theta":           "RoPE θ",
+    "unmask.flag.yes":             "oui",
+    "unmask.flag.no":              "non",
+    "unmask.flag.full_mha":        "non (MHA complet, {n} heads)",
+    "unmask.verdict.honest":            "✅ HONNÊTE",
+    "unmask.verdict.inflated":          "⚠ GONFLÉ",
+    "unmask.verdict.severely_inflated": "❌ GRAVEMENT GONFLÉ",
+    "unmask.verdict.yarn_extended":     "⚠ YARN-ÉTENDU",
+    "unmask.verdict.unknown":           "❓ INCONNU",
+    "unmask.warn.swa_window":      "Fenêtre SWA : {window} tokens — chaque couche n'attend que dans cette fenêtre.",
+    "unmask.warn.multihop":        "Estimation multi-hop : ~{multiHop} tokens (conservateur : fenêtre × {factor}).",
+    "unmask.warn.yarn":            "RoPE scaling ({type}) étend le contexte {factor}× de ~{original} à {declared} tokens.",
+    "unmask.warn.yarn_advice":     "Contexte RoPE-étendu — vérifiez le comportement de γ à la longueur déclarée avec le diagnostic γ_check.",
+    "unmask.warn.gqa_small_dhead": "Petite head dim ({d_head}) + GQA : compression de KV cache probable en contexte long (γ poussé vers Hagedorn).",
+    "unmask.reco.honest":              "Modèle d'attention complète standard. Contexte effectif correspond au déclaré ({declared} tokens).",
+    "unmask.reco.inflated":            "Effectif ~{effective} tokens via SWA. Utilisez γ_check pour vérifier le comportement à votre longueur cible.",
+    "unmask.reco.severely_inflated":   "Traitez-le comme un modèle de ~{effective} tokens en pratique. Le claim de {declared} tokens ne s'applique que via des chaînes d'attention cross-layer, qui dégradent empiriquement au-delà de ~2× la fenêtre SWA.",
+    "unmask.reco.yarn_extended":       "Contexte RoPE-étendu. Lancez un benchmark long-context (NIAH à 8k / 16k / 32k / full) pour confirmer que l'extension tient. Utilisez γ_check avec T_eval = {declared}.",
+    "unmask.reco.unknown":             "Impossible de parser le config. Vérifiez que l'URL est un modèle HF valide avec config.json public.",
+    "unmask.status.empty_id":      "⚠ Saisissez un model id (ex. mistralai/Mistral-7B-v0.1).",
+    "unmask.status.fetching":      "⏳ Récupération config.json pour {modelId}...",
+    "unmask.status.success":       "✅ {modelId} analysé (verdict : {verdict})",
+    "unmask.status.empty_paste":   "⚠ Collez d'abord un config.json.",
+    "unmask.status.invalid_json":  "❌ JSON invalide : {error}",
+    "unmask.status.success_paste": "✅ Config collé analysé (verdict : {verdict})",
+    "unmask.pasted_label":         "(config collé)",
+    "mode_desc.ask":               "Tapez une question libre. Le LLM dans le navigateur choisit la recette et l'exécute.",
+    "mode_desc.recipe":            "Sélectionnez une recette directement et remplissez le formulaire. Contrôle manuel complet.",
+    "mode_desc.profile":           "Démarrage le plus rapide : collez n'importe quel model id HuggingFace, cliquez Profile. Voyez les 5 recettes en quelques secondes.",
+    "mode_desc.compare":           "Choisissez 2-3 modèles candidats + une recette. Verdicts côte à côte dans un tableau.",
+    "mode_desc.inspector":         "Collez un config.json directement. Utile pour modèles privés / en dev non publiés sur HF Hub.",
+    "mode_desc.diagnose":          "Construit la commande CLI diagnose_model.py pour MESURER γ_obs sur GPU réel. Le navigateur prédit ; le CLI mesure.",
+    "mode_desc.phase":             "Scatter γ × θ du panel empirique du papier. Survolez les points pour détails, cliquez pour charger dans Diagnose / Recipe.",
+    "mode_desc.unmask":            "Détecte si max_position_embeddings est trompeur (SWA / YaRN / RoPE-scaling). Collez un model id, obtenez un verdict en 1 ligne.",
+    "profile.preset_loaded":       "✅ Préréglage chargé pour <strong>{id}</strong>. Formulaire pré-rempli. (Cliquez 📥 Fetch pour écraser avec le dernier config depuis HF Hub.)",
     "share.import_desc":       "Vous avez un fichier JSON de l'analyse TAF de quelqu'un ? Chargez-le ici pour voir le verdict + la chaîne localement. La même vue que si vous l'aviez exécuté vous-même.",
     "share.import_btn":        "📂 Charger JSON partagé",
     "synthesis.system":        "Vous êtes un assistant de diagnostic précis pour LLMs transformer. Étant donné des résultats de formules TAF pré-calculés, écrivez un résumé clair en français de 4-6 phrases. Citez le numéro de section (§X.Y) pour chaque nombre mentionné. Donnez toujours une recommandation concrète. N'INVENTEZ PAS de nombres.",
@@ -1203,7 +1371,7 @@ export const TRANSLATIONS = {
     "common.no":           "Non",
 
     // Tooltips des modes
-    "modes.tip":           "<strong>Sept façons d'utiliser l'outil</strong>.<br><strong>📇 Profil</strong>: collez un id → TAF Card avec 5 recettes.<br><strong>🆚 Comparer</strong>: 2-3 modèles côte à côte sur une recette.<br><strong>🔍 Inspecter config</strong>: collez config.json brut → Profil complet.<br><strong>💬 Question</strong>: question libre, le LLM du navigateur choisit la recette.<br><strong>📋 Recette</strong>: sélection manuelle avec contrôle total du formulaire.<br><strong>🩺 Diagnostic CLI</strong>: génère commande Python pour mesurer γ localement.<br><strong>📊 Diagramme de phase</strong>: panel de 23 modèles dans le plan (log θ, γ).",
+    "modes.tip":           "<strong>Huit façons d'utiliser l'outil</strong>.<br><strong>📇 Profil</strong>: collez un id → TAF Card avec 5 recettes.<br><strong>🆚 Comparer</strong>: 2-3 modèles côte à côte sur une recette.<br><strong>🔍 Inspecter config</strong>: collez config.json brut → Profil complet.<br><strong>💬 Question</strong>: question libre, le LLM du navigateur choisit la recette.<br><strong>📋 Recette</strong>: sélection manuelle avec contrôle total du formulaire.<br><strong>🩺 Diagnostic CLI</strong>: génère commande Python pour mesurer γ localement.<br><strong>📊 Diagramme de phase</strong>: panel de 23 modèles dans le plan (log θ, γ).<br><strong>🪟 Démasquer</strong>: détecte un max_position_embeddings trompeur (SWA / YaRN / RoPE-scaling).",
     "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.",
 
@@ -1524,7 +1692,7 @@ export const TRANSLATIONS = {
     "inv.export.formats":          "<strong>JSON · Markdown · LaTeX</strong>（论文级）",
     "inv.export.share":            "可复现的分享链接（状态编入 URL）",
     "inv.export.registry":         "提交到 GitHub 上的社区登记",
-    "arch.summary":                "支持的架构（点击展开）",
+    "arch.summary":                "支持的架构",
     "arch.anyhf":                  "✓ 任意 HuggingFace 公开模型",
     "tooltip.mha":                 "Multi-Head Attention：每个 token 位置同时通过多个并行 head 进行注意力计算。",
     "tooltip.gqa":                 "Grouped Query Attention：queries 共享比 heads 更少的 keys/values（节省内存但把 γ 推向 Hagedorn）。",
@@ -1532,6 +1700,62 @@ export const TRANSLATIONS = {
     "tooltip.abspe":               "Absolute Position Embeddings：每个位置有一个固定的学习向量加到 token embedding。",
     "tooltip.swa":                 "Sliding Window Attention：每个 token 仅在固定局部窗口内做注意力（Mistral、gemma-2 使用此机制）。",
     "tooltip.ssm":                 "State Space Model：维护内部状态的序列层（取代注意力，Mamba、Jamba 使用此机制）。",
+
+    // v0.7.0 — anti-bullshit pack #1: SWA / RoPE-scaling 揭示器
+    "modes.unmask":                "🪟 揭示",
+    "unmask.title":                "🪟 上下文揭示器",
+    "unmask.tip":                  "粘贴 HuggingFace 模型 id（或原始 config.json）。工具检测 sliding-window attention、RoPE 缩放（YaRN/linear/dynamic NTK）和 GQA — 所有使 <code>max_position_embeddings</code> 大于实际有效上下文的因素。Mistral-7B-v0.1 是经典例子：声称 32k，实际只在 ~4-8k 范围内做注意力。",
+    "unmask.desc":                 "<strong>你即将为一个实际上注意力不到那么远的模型花钱吗？</strong> 粘贴 id，1 秒内得知。无需 GPU，无需推理 — 只是对 config.json 做算术。",
+    "unmask.id_label":             "HF 模型 id：",
+    "unmask.fetch_btn":            "🔍 揭示",
+    "unmask.paste_summary":        "或粘贴原始 config.json（私有 / 在研模型）",
+    "unmask.paste_btn":            "🔍 揭示已粘贴的 config",
+    "unmask.label.declared":       "声明上下文",
+    "unmask.label.effective":      "有效（估计）",
+    "unmask.label.ratio":          "比率",
+    "unmask.section.flags":        "架构标志",
+    "unmask.section.warnings":     "警告",
+    "unmask.section.reco":         "建议",
+    "unmask.flag.swa":             "SWA",
+    "unmask.flag.rope":            "RoPE 缩放",
+    "unmask.flag.gqa":             "GQA",
+    "unmask.flag.layers":          "层数",
+    "unmask.flag.dhead":           "d_head",
+    "unmask.flag.theta":           "RoPE θ",
+    "unmask.flag.yes":             "是",
+    "unmask.flag.no":              "否",
+    "unmask.flag.full_mha":        "否（完整 MHA，{n} heads）",
+    "unmask.verdict.honest":            "✅ 诚实",
+    "unmask.verdict.inflated":          "⚠ 夸大",
+    "unmask.verdict.severely_inflated": "❌ 严重夸大",
+    "unmask.verdict.yarn_extended":     "⚠ YARN 扩展",
+    "unmask.verdict.unknown":           "❓ 未知",
+    "unmask.warn.swa_window":      "SWA 窗口：{window} tokens — 每层仅在此窗口内做注意力。",
+    "unmask.warn.multihop":        "多跳估计：~{multiHop} tokens（保守：窗口 × {factor}）。",
+    "unmask.warn.yarn":            "RoPE 缩放（{type}）将上下文从 ~{original} 扩展 {factor}× 到 {declared} tokens。",
+    "unmask.warn.yarn_advice":     "RoPE 扩展的上下文 — 用 γ_check 诊断在声称的全长度验证 γ 行为。",
+    "unmask.warn.gqa_small_dhead": "小 head dim（{d_head}）+ GQA：长上下文下 KV 缓存压缩很可能（γ 推向 Hagedorn）。",
+    "unmask.reco.honest":              "标准全注意力模型。有效上下文与声明一致（{declared} tokens）。",
+    "unmask.reco.inflated":            "通过 SWA 有效 ~{effective} tokens。用 γ_check 验证你目标长度的行为。",
+    "unmask.reco.severely_inflated":   "实际把它当作 ~{effective} tokens 上下文模型。{declared} tokens 的声明仅通过跨层注意力链生效，经验上超过 ~2× SWA 窗口后会退化。",
+    "unmask.reco.yarn_extended":       "RoPE 扩展上下文。运行长上下文 benchmark（NIAH 在 8k / 16k / 32k / 全长度）以确认扩展是否成立。用 γ_check 设 T_eval = {declared}。",
+    "unmask.reco.unknown":             "无法解析 config。验证 URL 是带公开 config.json 的有效 HF 模型。",
+    "unmask.status.empty_id":      "⚠ 输入一个 model id（例如 mistralai/Mistral-7B-v0.1）。",
+    "unmask.status.fetching":      "⏳ 正在获取 {modelId} 的 config.json...",
+    "unmask.status.success":       "✅ 已分析 {modelId}（判定：{verdict}）",
+    "unmask.status.empty_paste":   "⚠ 请先粘贴 config.json。",
+    "unmask.status.invalid_json":  "❌ JSON 无效：{error}",
+    "unmask.status.success_paste": "✅ 已分析粘贴的 config（判定：{verdict}）",
+    "unmask.pasted_label":         "（已粘贴 config）",
+    "mode_desc.ask":               "输入自由问题。浏览器内的 LLM 选择正确的 recipe 并运行。",
+    "mode_desc.recipe":            "直接选择一个 recipe 并填表。完整手动控制。",
+    "mode_desc.profile":           "最快开始：粘贴任意 HuggingFace model id，点击 Profile。几秒内看到 5 个 recipe。",
+    "mode_desc.compare":           "选择 2-3 个候选模型 + 一个 recipe。在表格中并排查看判定。",
+    "mode_desc.inspector":         "直接粘贴 config.json。适用于未发布 HF Hub 的私有 / 在研模型。",
+    "mode_desc.diagnose":          "构建 diagnose_model.py 的 CLI 命令，在真实 GPU 上测量 γ_obs。浏览器预测；CLI 测量。",
+    "mode_desc.phase":             "论文经验面板的 γ × θ 散点图。悬停点查看详情，点击加载到 Diagnose / Recipe 表单。",
+    "mode_desc.unmask":            "检测 max_position_embeddings 是否误导（SWA / YaRN / RoPE 缩放）。粘贴 model id，1 行判定。",
+    "profile.preset_loaded":       "✅ 已为 <strong>{id}</strong> 加载预设。表单已预填。（点击 📥 Fetch 用 HF Hub 最新 config 覆盖。）",
     "share.import_desc":       "有他人 TAF 分析的 JSON 文件? 在这里加载以本地查看判定 + 链。与您自己运行的视图相同。",
     "share.import_btn":        "📂 加载共享的 JSON",
     "synthesis.system":        "您是 transformer LLM 的精确诊断助手。给定预先计算的 TAF 公式结果,用 4-6 句中文写出清晰的摘要。为每个提到的数字引用章节号 (§X.Y)。始终给出具体建议。不要编造数字。",
@@ -1624,7 +1848,7 @@ export const TRANSLATIONS = {
     "common.no":           "否",
 
     // 模式提示
-    "modes.tip":           "<strong>七种使用方式</strong>。<br><strong>📇 画像</strong>: 粘贴模型 id → 5 个配方的 TAF 卡。<br><strong>🆚 比较</strong>: 2-3 个模型在一个配方上并排比较。<br><strong>🔍 检查 config</strong>: 粘贴原始 config.json → 完整画像。<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。<br><strong>🩺 CLI 诊断</strong>: 生成 Python 命令在本地测量 γ。<br><strong>📊 相图</strong>: 23 个面板模型在 (log θ, γ) 平面上。",
+    "modes.tip":           "<strong>八种使用方式</strong>。<br><strong>📇 画像</strong>: 粘贴模型 id → 5 个配方的 TAF 卡。<br><strong>🆚 比较</strong>: 2-3 个模型在一个配方上并排比较。<br><strong>🔍 检查 config</strong>: 粘贴原始 config.json → 完整画像。<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。<br><strong>🩺 CLI 诊断</strong>: 生成 Python 命令在本地测量 γ。<br><strong>📊 相图</strong>: 23 个面板模型在 (log θ, γ) 平面上。<br><strong>🪟 揭示</strong>: 检测误导的 max_position_embeddings（SWA / YaRN / RoPE 缩放）。",
     "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。",
 

js/main.js

@@ -11,6 +11,7 @@ import { initI18n, setLang, t } from "./i18n.js";
 import { initPhaseDiagram } from "./phase_diagram.js";
 import { gammaCheckAll, REGIME_META } from "./gamma_check.js";
 import { loadLeanManifest, badgeHtml, badgesForUiBinding, renderTheoremTable, getManifest } from "./lean_badges.js";
+import { unmaskConfig } from "./swa_unmasker.js";
 
 const TAF_BROWSER_URL = "python/taf_browser.py";
 const ENABLE_WEBLLM = true;
@@ -137,6 +138,38 @@ function enableUI() {
 
 function setStatus(msg) { $("status").textContent = msg; }
 
+// ════════════════════════════════════════════════════════════════════
+// Main-panel wrap: every <main> section gets a foldable details/summary
+// shell at runtime so users can collapse any panel they don't need open.
+// h2 is moved INTO summary so its data-i18n binding survives. Idempotent.
+// ════════════════════════════════════════════════════════════════════
+function wrapMainSectionsAsFoldable() {
+  document.querySelectorAll("main > section").forEach(section => {
+    if (section.id === "status-bar") return;                     // skip loading bar
+    if (section.querySelector(":scope > details.main-panel")) return; // already wrapped
+    const h2 = section.querySelector(":scope > h2");
+    if (!h2) return;
+
+    const details = document.createElement("details");
+    details.className = "main-panel";
+    details.open = true;
+
+    const summary = document.createElement("summary");
+    summary.className = "main-panel-title";
+    summary.appendChild(h2);  // preserve h2 + its data-i18n + all children
+
+    details.appendChild(summary);
+    while (section.firstChild) details.appendChild(section.firstChild);
+    section.appendChild(details);
+  });
+
+  // Stop ⓘ tooltip clicks inside summaries from toggling the panel.
+  document.querySelectorAll(".main-panel > .main-panel-title .info").forEach(el => {
+    el.addEventListener("click", (e) => e.stopPropagation());
+  });
+}
+wrapMainSectionsAsFoldable();
+
 // ════════════════════════════════════════════════════════════════════
 // Mode toggle
 // ════════════════════════════════════════════════════════════════════
@@ -153,41 +186,20 @@ document.querySelectorAll(".mode-btn").forEach(btn => {
     // Hide all mode sections
     ["ask-section", "recipe-section", "form-section",
      "profile-section", "compare-section", "inspector-section",
-     "diagnose-section", "phase-section"].forEach(id => {
+     "diagnose-section", "phase-section", "unmask-section"].forEach(id => {
       const el = $(id);
       if (el) el.style.display = "none";
     });
     // Show selected
-    if (mode === "ask") {
-      $("ask-section").style.display = "";
-      $("mode-desc").textContent =
-        "Type a free-form question. The in-browser LLM picks the right recipe and runs it.";
-    } else if (mode === "recipe") {
-      $("recipe-section").style.display = "";
-      $("mode-desc").textContent =
-        "Pick a recipe directly and fill the form. Full manual control.";
-    } else if (mode === "profile") {
-      $("profile-section").style.display = "";
-      $("mode-desc").textContent =
-        "Quickest start: paste any HuggingFace model id, click Profile. See all 5 recipes scored in seconds.";
-    } else if (mode === "compare") {
-      $("compare-section").style.display = "";
-      $("mode-desc").textContent =
-        "Pick 2-3 candidate models + one recipe. See verdicts side-by-side in a comparison table.";
-    } else if (mode === "inspector") {
-      $("inspector-section").style.display = "";
-      $("mode-desc").textContent =
-        "Paste a config.json directly. Useful for private/in-development models not on HF Hub.";
-    } else if (mode === "diagnose") {
-      $("diagnose-section").style.display = "";
-      $("mode-desc").textContent =
-        "Build the diagnose_model.py CLI command to MEASURE γ_obs on real GPU. Browser predicts; CLI measures.";
-    } else if (mode === "phase") {
-      $("phase-section").style.display = "";
-      $("mode-desc").textContent =
-        "γ × θ scatter of the paper's empirical panel. Hover a dot for details, click to load into Diagnose / Recipe forms.";
-      initPhaseDiagram();
-    }
+    const sectionMap = {
+      ask: "ask-section", recipe: "recipe-section", profile: "profile-section",
+      compare: "compare-section", inspector: "inspector-section",
+      diagnose: "diagnose-section", phase: "phase-section", unmask: "unmask-section",
+    };
+    const sectionId = sectionMap[mode];
+    if (sectionId) $(sectionId).style.display = "";
+    $("mode-desc").textContent = t(`mode_desc.${mode}`) || "";
+    if (mode === "phase") initPhaseDiagram();
   });
 });
 
@@ -353,8 +365,14 @@ function getRecipeDefaults(recipeId) {
 // ════════════════════════════════════════════════════════════════════
 $("preset").addEventListener("change", (e) => {
   if (!e.target.value) return;
-  state.lastModelId = e.target.value;  // remember for filename/hash
-  const proxy = state.pyodide.runPython(`get_preset(${JSON.stringify(e.target.value)})`);
+  const modelId = e.target.value;
+  state.lastModelId = modelId;  // remember for filename/hash
+  // Mirror behavior with profile-preset: also fill HF id input if present.
+  if ($("hf-id")) {
+    $("hf-id").value = modelId;
+    if ($("hf-status")) $("hf-status").textContent = tFmt("profile.preset_loaded", { id: modelId });
+  }
+  const proxy = state.pyodide.runPython(`get_preset(${JSON.stringify(modelId)})`);
   const preset = proxy.toJs ? proxy.toJs({ dict_converter: Object.fromEntries }) : proxy;
   if (!preset || Object.keys(preset).length === 0) return;
   fillRecipeForm(preset);
@@ -417,6 +435,152 @@ $("hf-fetch-btn").addEventListener("click", async () => {
   }
 });
 
+// ════════════════════════════════════════════════════════════════════
+// 🪟 Unmask mode (v0.7.0 anti-bullshit pack #1)
+// ════════════════════════════════════════════════════════════════════
+
+// Tiny string-template helper: t(key) with {placeholder} substitution.
+// Falls back to the raw key when the i18n entry is missing so dev sees the gap.
+function tFmt(key, params = {}) {
+  let s = t(key) || key;
+  for (const [k, v] of Object.entries(params)) {
+    const fmtVal = v === null || v === undefined ? "—"
+      : (typeof v === "number" ? v.toLocaleString() : String(v));
+    s = s.replace(new RegExp(`\\{${k}\\}`, "g"), fmtVal);
+  }
+  return s;
+}
+
+const VERDICT_COLOR = {
+  honest:            "#3fb950",
+  inflated:          "#f1c40f",
+  severely_inflated: "#f85149",
+  yarn_extended:     "#f1c40f",
+  unknown:           "#8b949e",
+};
+
+function renderUnmaskCard(result, modelId = "") {
+  const color = VERDICT_COLOR[result.verdict] || VERDICT_COLOR.unknown;
+  const ratioPct = (result.ratio * 100).toFixed(1);
+  const f = result.flags;
+  const fmtN = (x) => x === null || x === undefined ? "—" : Number(x).toLocaleString();
+  const escapeHtml = (s) => String(s).replace(/[&<>"']/g, c =>
+    ({"&":"&amp;","<":"&lt;",">":"&gt;",'"':"&quot;","'":"&#39;"}[c]));
+
+  const verdictLabel = t(`unmask.verdict.${result.verdict}`) || result.verdict;
+  const labelDeclared  = t("unmask.label.declared")  || "Declared context";
+  const labelEffective = t("unmask.label.effective") || "Effective (estimate)";
+  const labelRatio     = t("unmask.label.ratio")     || "Ratio";
+  const sectionFlags   = t("unmask.section.flags")   || "Architecture flags";
+  const sectionWarn    = t("unmask.section.warnings")|| "Warnings";
+  const sectionReco    = t("unmask.section.reco")    || "Recommendation";
+
+  // Architecture flags row labels
+  const flagSwa     = t("unmask.flag.swa")     || "SWA";
+  const flagRope    = t("unmask.flag.rope")    || "RoPE scaling";
+  const flagGqa     = t("unmask.flag.gqa")     || "GQA";
+  const flagLayers  = t("unmask.flag.layers")  || "Layers";
+  const flagDhead   = t("unmask.flag.dhead")   || "d_head";
+  const flagTheta   = t("unmask.flag.theta")   || "RoPE θ";
+  const flagYes     = t("unmask.flag.yes")     || "yes";
+  const flagNo      = t("unmask.flag.no")      || "no";
+
+  const swaText = f.hasSWA
+    ? `${flagYes} (window = ${fmtN(f.swaWindow)})`
+    : flagNo;
+  const ropeText = f.hasYaRN
+    ? `${f.ropeScalingType} (factor = ${f.yarnFactor}, original = ${fmtN(f.yarnOriginal)})`
+    : flagNo;
+  const gqaText = f.hasGQA
+    ? `${flagYes} (${f.n_kv_heads} kv / ${f.n_attn_heads} attn heads)`
+    : (t("unmask.flag.full_mha") || "no (full MHA, {n} heads)").replace("{n}", f.n_attn_heads ?? "?");
+
+  const warningsHtml = result.warnings.length
+    ? `<details class="unmask-panel" open><summary class="unmask-panel-title">${sectionWarn}</summary><ul>${result.warnings.map(w =>
+        `<li>${tFmt("unmask.warn." + w.code, w.params)}</li>`).join("")}</ul></details>`
+    : "";
+
+  const recoHtml = result.recoCode
+    ? `<details class="unmask-panel" open><summary class="unmask-panel-title">${sectionReco}</summary><p class="unmask-reco">${tFmt("unmask.reco." + result.recoCode, result.recoParams)}</p></details>`
+    : "";
+
+  return `
+    <div class="unmask-result">
+      <div class="unmask-hero" style="border-color: ${color};">
+        <div class="unmask-verdict" style="color: ${color};">${verdictLabel}</div>
+        ${modelId ? `<div class="unmask-model"><code>${escapeHtml(modelId)}</code></div>` : ""}
+        <div class="unmask-numbers">
+          <div><span class="unmask-num-label">${labelDeclared}</span><span class="unmask-num-val">${fmtN(result.declaredContext)}</span></div>
+          <div><span class="unmask-num-label">${labelEffective}</span><span class="unmask-num-val">${fmtN(result.effectiveContext)}</span></div>
+          <div><span class="unmask-num-label">${labelRatio}</span><span class="unmask-num-val">${ratioPct}%</span></div>
+        </div>
+      </div>
+
+      <div class="unmask-details">
+        <details class="unmask-panel" open>
+          <summary class="unmask-panel-title">${sectionFlags}</summary>
+          <ul>
+            <li><strong>${flagSwa}:</strong> ${swaText}</li>
+            <li><strong>${flagRope}:</strong> ${ropeText}</li>
+            <li><strong>${flagGqa}:</strong> ${gqaText}</li>
+            <li><strong>${flagLayers}:</strong> ${fmtN(f.n_layers)} · <strong>${flagDhead}:</strong> ${fmtN(f.d_head)} · <strong>${flagTheta}:</strong> ${fmtN(f.rope_theta)}</li>
+          </ul>
+        </details>
+        ${warningsHtml}
+        ${recoHtml}
+      </div>
+    </div>
+  `;
+}
+
+async function runUnmaskFromId() {
+  const modelId = ($("unmask-id").value || "").trim();
+  if (!modelId) {
+    $("unmask-status").textContent = t("unmask.status.empty_id") || "⚠ Enter a model id.";
+    return;
+  }
+  $("unmask-status").textContent = tFmt("unmask.status.fetching", { modelId });
+  $("unmask-fetch-btn").disabled = true;
+  try {
+    const cfg = await fetchHfConfig(modelId);
+    const result = unmaskConfig(cfg);
+    $("unmask-output").innerHTML = renderUnmaskCard(result, modelId);
+    const verdictLocalized = t(`unmask.verdict.${result.verdict}`) || result.verdict;
+    $("unmask-status").textContent = tFmt("unmask.status.success", { modelId, verdict: verdictLocalized });
+  } catch (err) {
+    $("unmask-status").textContent = `❌ ${err.message}`;
+    $("unmask-output").innerHTML = "";
+  } finally {
+    $("unmask-fetch-btn").disabled = false;
+  }
+}
+
+function runUnmaskFromPaste() {
+  const raw = ($("unmask-paste").value || "").trim();
+  if (!raw) {
+    $("unmask-status").textContent = t("unmask.status.empty_paste") || "⚠ Paste a config.json first.";
+    return;
+  }
+  let cfg;
+  try {
+    cfg = JSON.parse(raw);
+  } catch (e) {
+    $("unmask-status").textContent = tFmt("unmask.status.invalid_json", { error: e.message });
+    return;
+  }
+  const result = unmaskConfig(cfg);
+  const pastedLabel = t("unmask.pasted_label") || "(pasted config)";
+  $("unmask-output").innerHTML = renderUnmaskCard(result, pastedLabel);
+  const verdictLocalized = t(`unmask.verdict.${result.verdict}`) || result.verdict;
+  $("unmask-status").textContent = tFmt("unmask.status.success_paste", { verdict: verdictLocalized });
+}
+
+$("unmask-fetch-btn")?.addEventListener("click", runUnmaskFromId);
+$("unmask-paste-btn")?.addEventListener("click", runUnmaskFromPaste);
+$("unmask-id")?.addEventListener("keydown", (e) => {
+  if (e.key === "Enter") { e.preventDefault(); runUnmaskFromId(); }
+});
+
 function configToPreset(cfg, modelId) {
   const n_attn = cfg.num_attention_heads || cfg.n_head || 0;
   const n_kv = cfg.num_key_value_heads || cfg.num_attention_heads || cfg.n_head || 0;
@@ -988,8 +1152,18 @@ function relativeTime(d) {
 // ════════════════════════════════════════════════════════════════════
 $("profile-preset").addEventListener("change", (e) => {
   if (!e.target.value) return;
-  state.lastModelId = e.target.value;  // remember for filename/hash
-  const proxy = state.pyodide.runPython(`get_preset(${JSON.stringify(e.target.value)})`);
+  const modelId = e.target.value;
+  state.lastModelId = modelId;  // remember for filename/hash
+  // Preset keys ARE valid HF model ids (e.g. "meta-llama/Llama-3.2-1B"). Auto-fill
+  // the HF id input so the user can also click 📥 Fetch to refresh from HF Hub
+  // without retyping. Status hint clarifies the dual source of truth.
+  if ($("profile-hf-id")) {
+    $("profile-hf-id").value = modelId;
+    if ($("profile-hf-status")) {
+      $("profile-hf-status").textContent = tFmt("profile.preset_loaded", { id: modelId });
+    }
+  }
+  const proxy = state.pyodide.runPython(`get_preset(${JSON.stringify(modelId)})`);
   const p = proxy.toJs ? proxy.toJs({ dict_converter: Object.fromEntries }) : proxy;
   if (!p || Object.keys(p).length === 0) return;
   $("profile-theta").value = p.theta;

js/swa_unmasker.js

@@ -0,0 +1,107 @@
+// SWA Unmasker (v0.7.0 anti-bullshit pack #1)
+// Pure logic — no human-readable strings. Returns structured warnings/reco
+// codes + params; main.js does the i18n lookup so EN/ES/FR/ZH all work.
+
+// Conservative multi-hop bound for SWA models. Empirically the effective
+// "reasoning" context is roughly 2× the window, NOT window × n_layers
+// (which is the theoretical upper bound but breaks down past a few hops).
+const SWA_MULTIHOP_FACTOR = 2;
+
+export function unmaskConfig(config) {
+  const out = {
+    declaredContext: config.max_position_embeddings ?? null,
+    effectiveContext: null,
+    verdict: "honest",
+    ratio: 1.0,
+    flags: {
+      hasSWA: false,
+      swaWindow: null,
+      hasYaRN: false,
+      yarnFactor: null,
+      yarnOriginal: null,
+      ropeScalingType: null,
+      hasGQA: false,
+      n_kv_heads: config.num_key_value_heads ?? config.num_attention_heads ?? null,
+      n_attn_heads: config.num_attention_heads ?? null,
+      n_layers: config.num_hidden_layers ?? null,
+      rope_theta: config.rope_theta ?? null,
+      d_head: null,
+    },
+    warnings: [],   // each: { code, params }
+    recoCode: null,
+    recoParams: {},
+  };
+
+  if (out.flags.n_attn_heads && out.flags.n_kv_heads) {
+    out.flags.hasGQA = out.flags.n_kv_heads < out.flags.n_attn_heads;
+  }
+  if (config.hidden_size && out.flags.n_attn_heads) {
+    out.flags.d_head = config.hidden_size / out.flags.n_attn_heads;
+  }
+
+  // SWA: explicit sliding_window field (Mistral, Gemma-2). Some configs set
+  // it to null or to max_pe — treat as "no SWA" in those cases.
+  const sw = config.sliding_window;
+  if (typeof sw === "number" && sw > 0
+      && (!out.declaredContext || sw < out.declaredContext)) {
+    out.flags.hasSWA = true;
+    out.flags.swaWindow = sw;
+  }
+
+  // RoPE scaling (YaRN / linear / dynamic NTK). Only flag if factor > 1.
+  const rs = config.rope_scaling;
+  if (rs && typeof rs === "object") {
+    out.flags.ropeScalingType = rs.type ?? rs.rope_type ?? null;
+    out.flags.yarnFactor = rs.factor ?? null;
+    out.flags.yarnOriginal = rs.original_max_position_embeddings ?? null;
+    if (out.flags.ropeScalingType && out.flags.yarnFactor && out.flags.yarnFactor > 1) {
+      out.flags.hasYaRN = true;
+    }
+  }
+
+  // Compute verdict
+  if (out.flags.hasSWA) {
+    const multiHop = out.flags.swaWindow * SWA_MULTIHOP_FACTOR;
+    out.effectiveContext = Math.min(multiHop, out.declaredContext ?? multiHop);
+    out.ratio = out.declaredContext ? out.effectiveContext / out.declaredContext : 1.0;
+    // <= 0.25 catches the canonical Mistral case (window=4096, declared=32768, ratio=0.25 exact)
+    out.verdict = out.ratio <= 0.25 ? "severely_inflated" : "inflated";
+    out.warnings.push(
+      { code: "swa_window", params: { window: out.flags.swaWindow } },
+      { code: "multihop", params: { multiHop, factor: SWA_MULTIHOP_FACTOR } },
+    );
+    out.recoCode = out.verdict;
+    out.recoParams = {
+      effective: out.effectiveContext,
+      declared: out.declaredContext,
+    };
+  } else if (out.flags.hasYaRN) {
+    out.verdict = "yarn_extended";
+    const orig = out.flags.yarnOriginal
+      ?? (out.declaredContext ? out.declaredContext / out.flags.yarnFactor : null);
+    out.effectiveContext = out.declaredContext;
+    out.ratio = 1.0;
+    out.warnings.push(
+      { code: "yarn", params: { type: out.flags.ropeScalingType, factor: out.flags.yarnFactor, original: orig ? Math.round(orig) : null, declared: out.declaredContext } },
+      { code: "yarn_advice", params: {} },
+    );
+    out.recoCode = "yarn_extended";
+    out.recoParams = { declared: out.declaredContext };
+  } else if (out.declaredContext) {
+    out.effectiveContext = out.declaredContext;
+    out.verdict = "honest";
+    out.recoCode = "honest";
+    out.recoParams = { declared: out.declaredContext };
+  } else {
+    out.verdict = "unknown";
+    out.recoCode = "unknown";
+    out.recoParams = {};
+  }
+
+  // KV-cache compression hint for small d_head + GQA — independent of verdict
+  if (out.flags.hasGQA && out.flags.d_head && out.flags.d_head < 64) {
+    out.warnings.push({ code: "gqa_small_dhead", params: { d_head: out.flags.d_head } });
+  }
+
+  return out;
+}

style.css

@@ -1,5 +1,130 @@
 /* TAF Agent — minimal clean styling */
 
+/* v0.7.0 — main panels foldable (every section under <main>) */
+.main-panel { margin: 0; }
+.main-panel > .main-panel-title {
+  cursor: pointer;
+  list-style: none;
+  user-select: none;
+  padding: 0 0 0.5em;
+  margin-bottom: 0.6em;
+  border-bottom: 1px solid rgba(255, 255, 255, 0.06);
+  display: flex;
+  align-items: baseline;
+  gap: 0.5em;
+}
+.main-panel > .main-panel-title::-webkit-details-marker { display: none; }
+.main-panel > .main-panel-title::marker { content: ""; }
+.main-panel > .main-panel-title::before {
+  content: "▼";
+  display: inline-block;
+  font-size: 0.65em;
+  color: #58a6ff;
+  margin-right: 0.3em;
+  transition: transform 0.15s ease;
+  flex-shrink: 0;
+}
+.main-panel:not([open]) > .main-panel-title::before { transform: rotate(-90deg); }
+.main-panel > .main-panel-title:hover { background: rgba(255, 255, 255, 0.02); }
+.main-panel > .main-panel-title h2 {
+  display: inline;
+  margin: 0;
+  vertical-align: baseline;
+  flex: 1;
+}
+
+/* v0.7.0 — Unmask mode (SWA + RoPE-scaling detector) */
+.unmask-result {
+  margin-top: 0.8em;
+}
+.unmask-hero {
+  padding: 1em 1.2em;
+  border: 2px solid #58a6ff;
+  border-radius: 10px;
+  background: #12181f;
+  margin-bottom: 0.8em;
+}
+.unmask-verdict {
+  font-size: 1.6em;
+  font-weight: 700;
+  margin-bottom: 0.2em;
+}
+.unmask-model {
+  font-size: 0.92em;
+  opacity: 0.85;
+  margin-bottom: 0.6em;
+}
+.unmask-numbers {
+  display: grid;
+  grid-template-columns: repeat(auto-fit, minmax(140px, 1fr));
+  gap: 0.6em;
+  margin-top: 0.5em;
+}
+.unmask-numbers > div {
+  display: flex;
+  flex-direction: column;
+  padding: 0.5em 0.7em;
+  background: rgba(0, 0, 0, 0.25);
+  border-radius: 6px;
+}
+.unmask-num-label {
+  font-size: 0.78em;
+  opacity: 0.75;
+  text-transform: uppercase;
+  letter-spacing: 0.04em;
+}
+.unmask-num-val {
+  font-size: 1.3em;
+  font-weight: 600;
+  font-family: monospace;
+  margin-top: 0.15em;
+}
+.unmask-details {
+  padding: 0.8em 1em;
+  background: #12181f;
+  border: 1px solid rgba(255, 255, 255, 0.08);
+  border-radius: 8px;
+}
+.unmask-details h4,
+.unmask-panel-title {
+  margin: 0.4em 0 0.3em;
+  color: #58a6ff;
+  font-size: 0.95em;
+  cursor: pointer;
+  list-style: none;
+  user-select: none;
+  font-weight: 600;
+}
+.unmask-panel-title::-webkit-details-marker { display: none; }
+.unmask-panel-title::marker { content: ""; }
+.unmask-panel-title::before {
+  content: "▼";
+  display: inline-block;
+  font-size: 0.75em;
+  margin-right: 0.4em;
+  color: #58a6ff;
+  transition: transform 0.15s ease;
+  width: 0.9em;
+  text-align: center;
+}
+.unmask-panel:not([open]) > .unmask-panel-title::before { transform: rotate(-90deg); }
+.unmask-panel { margin: 0.5em 0; }
+.unmask-details ul {
+  margin: 0.2em 0 0.6em;
+  padding-left: 1.2em;
+  font-size: 0.92em;
+  line-height: 1.5;
+}
+.unmask-reco {
+  margin: 0.2em 0 0.4em;
+  padding: 0.6em 0.8em;
+  background: rgba(88, 166, 255, 0.08);
+  border-left: 3px solid #58a6ff;
+  border-radius: 0 6px 6px 0;
+  font-size: 0.92em;
+  line-height: 1.5;
+}
+
 /* v0.6.2 — landing rework: quick-start strip + inventory grid + arch-supported */
 #quickstart-strip {
   margin: 1.5em auto 1em;
@@ -72,18 +197,35 @@
   gap: 0.8em;
 }
 .inv-card {
-  padding: 0.9em 1em;
+  padding: 0.7em 1em;
   background: #12181f;
   border: 1px solid rgba(255, 255, 255, 0.08);
   border-radius: 8px;
 }
-.inv-card h3 {
-  margin: 0 0 0.5em;
+.inv-card-title {
+  cursor: pointer;
   font-size: 1em;
+  font-weight: 600;
   color: #58a6ff;
+  padding: 0.2em 0;
+  list-style: none;       /* hide native marker (Chrome, Safari) */
+  user-select: none;
+}
+.inv-card-title::-webkit-details-marker { display: none; } /* Safari */
+.inv-card-title::marker { content: ""; }                   /* Firefox */
+.inv-card-title::before {
+  content: "▼";
+  display: inline-block;
+  font-size: 0.75em;
+  margin-right: 0.4em;
+  color: #58a6ff;
+  transition: transform 0.15s ease;
+  width: 0.9em;
+  text-align: center;
 }
+.inv-card:not([open]) > .inv-card-title::before { transform: rotate(-90deg); }
 .inv-card ul {
-  margin: 0;
+  margin: 0.4em 0 0;
   padding-left: 1em;
   font-size: 0.92em;
   line-height: 1.5;