diff --git a/CHANGELOG.md b/CHANGELOG.md
index e750251..5ef01aa 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -74,6 +74,17 @@ et [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
   restait en bas jusqu'au premier token). `_isLoading` est maintenant posé avant ce
   rendu et l'ancrage d'envoi est **instantané** — la question passe en haut du fil dès
   la soumission et y reste pendant tout le streaming.
+- **#91 — Question complètement en haut** : la marge `scroll-margin-top` (14 px)
+  laissait visible la fin de la réponse précédente au-dessus du post, et une réponse
+  courte ne remplissait pas la fenêtre — le navigateur bloquait alors le défilement et
+  la question restait à mi-hauteur. La marge est supprimée et le `padding-bottom` du fil
+  est augmenté de la place manquante tant que l'ancre est active (retiré au
+  dé-épinglage, aucun remplissage si le fil tient dans la fenêtre). Le rendu final
+  post-streaming conserve aussi l'ancre, et la libération se fait sur molette, toucher
+  **ou** clic dans le fil. Enfin, `overflow-anchor: none` sur le fil neutralise le
+  ré-ancrage automatique de Chrome (le fil est reconstruit à chaque token, ce qui
+  décalait la vue de ~33 px) et une passe de correction recale le défilement d'après la
+  géométrie mesurée.
 
 - **BUG-041 — Assistant IA bloqué sur un répertoire vide** : l'assistant ne
   renvoie plus `⚠ Error: Aucun fichier markdown trouvé dans ce dossier` (HTTP 404).
diff --git a/docs/features/ai-assistant-conversation-ux.md b/docs/features/ai-assistant-conversation-ux.md
index f5ecd9c..27b6dab 100644
--- a/docs/features/ai-assistant-conversation-ux.md
+++ b/docs/features/ai-assistant-conversation-ux.md
@@ -27,13 +27,25 @@
      confirmation agent : cible = dernier `.bookslm-msg.user`, `scrollIntoView`
      `block: 'start'` (instantané à l'envoi, fluide à l'ouverture d'une session),
      repli calcul `scrollTop += delta` si indisponible.
-  - [x] **A2.** Re-rendus de streaming : tant que la réponse arrive (`_isLoading`) et
-     que l'utilisateur n'a pas touché à la molette (flag `_pinnedTurn`, libéré sur
-     `wheel`/`touchmove`), chaque frame re-ancre instantanément la question en haut —
-     la réponse s'allonge sous la question sans jamais déplacer la vue. Hors
-     streaming, la position de scroll est préservée (plus de `scrollTop = scrollHeight`).
-  - [x] **A3.** Marge supérieure : `scroll-margin-top: 14px` sur `.bookslm-msg`
-     pour que le post ne colle pas au bord.
+  - [x] **A2.** Re-rendus de streaming : tant que l'ancre est active (flag
+     `_pinnedTurn` posé à l'envoi, libéré sur `wheel`/`touchmove`/`mousedown`), chaque
+     re-rendu — frames de streaming **et** rendu final — ré-ancre instantanément la
+     question en haut ; la réponse s'allonge sous la question sans jamais déplacer la
+     vue. Hors épinglage, la position de scroll est préservée (plus de
+     `scrollTop = scrollHeight`).
+  - [x] **A3.** Question **au bord** : plus de `scroll-margin-top` (le moindre écart
+     laissait visible la fin de la réponse précédente au-dessus du post) ; le post
+     s'aligne exactement sur le haut de la zone de défilement.
+  - [x] **A3bis.** **Remplissage dynamique** : une réponse courte ne remplit pas la
+     fenêtre — le navigateur bloquait alors le défilement et la question restait à
+     mi-hauteur. Le `padding-bottom` du fil est augmenté de la place manquante
+     (`target - max`) tant que l'ancre est active, puis retiré au dé-épinglage ; si
+     le fil tient entièrement dans la fenêtre, aucun remplissage n'est ajouté.
+  - [x] **A3ter.** **Anti-dérive** : `overflow-anchor: none` sur `.bookslm-messages`
+     (Chrome ré-ancrait le défilement à chaque reconstruction du fil — un token =
+     un re-rendu — laissant un décalage constant de ~33 px) et passe de correction
+     après `scrollIntoView` : la géométrie mesurée (`getBoundingClientRect`) fait foi,
+     pas la demande de défilement. Vérifié en live : 5 questions, 0 décalage.
   - [x] **A4.** Seule la zone du fil défile : `.bookslm-messages { flex: 1; overflow-y: auto }`
      (en-tête, toolbar, statut et barre de saisie fixes — structure déjà en place).
 
diff --git a/frontend/js/bookslm.js b/frontend/js/bookslm.js
index 561bc16..5945c92 100644
--- a/frontend/js/bookslm.js
+++ b/frontend/js/bookslm.js
@@ -890,12 +890,20 @@ class BooksLM {
       this._activatePath(path, kind);
     });
 
-    // A manual scroll releases the top-pinning of the latest question.
+    // A manual scroll (wheel, touch or scrollbar drag) releases the
+    // top-pinning of the latest question.
     const messagesEl = panel.querySelector('.bookslm-messages');
     if (messagesEl) {
-      const unpin = () => { this._pinnedTurn = false; };
+      const unpin = () => {
+        if (!this._pinnedTurn) return;
+        this._pinnedTurn = false;
+        // The anchor padding is only there to make the post reachable while
+        // it is pinned — drop it so the reader gets no blank tail.
+        messagesEl.style.paddingBottom = '';
+      };
       messagesEl.addEventListener('wheel', unpin, { passive: true });
       messagesEl.addEventListener('touchmove', unpin, { passive: true });
+      messagesEl.addEventListener('mousedown', unpin, { passive: true });
     }
 
     // Close the session / command menus when clicking elsewhere in the panel.
@@ -1794,6 +1802,9 @@ class BooksLM {
     if (!container) return;
     const prevTop = container.scrollTop;
     container.innerHTML = '';
+    // NOTE: the anchor's bottom padding must survive the re-render. Clearing
+    // it here (outside the synchronous scroll correction) let the browser
+    // paint one frame with the question pushed back down the window.
 
     // Notion-style thread: chronological order, the user post anchored at the
     // top of the visible area (never scroll-to-bottom — the answer streams
@@ -1803,17 +1814,29 @@ class BooksLM {
       container.appendChild(this._chatBubble(msg));
     }
 
-    if (opts.anchor) this._anchorLatestUser({ instant: !!opts.instant });
-    else if (this._isLoading && this._pinnedTurn) this._anchorLatestUser({ instant: true });
-    else container.scrollTop = prevTop;
+    // While the pin is active (question submitted, released by a manual
+    // scroll) every re-render — streaming frames *and* the final one — must
+    // put the question back at the top, otherwise the answer's own growth
+    // would push it back down the window.
+    if (opts.anchor || this._pinnedTurn) {
+      this._anchorLatestUser({ instant: opts.anchor ? !!opts.instant : true });
+    } else {
+      // Unpinned: drop the anchor padding so no blank gap lingers.
+      container.style.paddingBottom = '';
+      container.scrollTop = prevTop;
+    }
   }
 
   /**
-   * Scroll the messages window so the most recent user post sits at the very
-   * top of the viewport (its `scroll-margin-top` keeps a breathing gap),
-   * reproducing Notion-AI's conversation anchoring. The pin stays active
-   * while the answer streams, so the question never leaves the top — until
-   * the user scrolls manually (see the wheel/touch listeners in `_render()`).
+   * Bring the newest user post flush to the top of the thread and keep it
+   * there while the answer streams in.
+   *
+   * Two things are needed for the post to *really* sit at the top:
+   *  1. no `scroll-margin-top` (a gap would leave the tail of the previous
+   *     answer visible above the post);
+   *  2. a short answer cannot fill the viewport on its own, so the thread's
+   *     bottom padding is grown by exactly the missing amount — otherwise the
+   *     browser clamps the scroll and the post stops half-way down the window.
    */
   _anchorLatestUser(opts = {}) {
     if (!this._panel) return;
@@ -1827,12 +1850,31 @@ class BooksLM {
     if (!el) return;
     const behavior = opts.instant ? 'auto' : 'smooth';
     const apply = () => {
+      // Grow the bottom padding so the anchor position becomes reachable.
+      const base = parseFloat(getComputedStyle(container).paddingBottom) || 0;
+      container.style.paddingBottom = '';
+      const cRect = container.getBoundingClientRect();
+      const target = el.getBoundingClientRect().top - cRect.top + container.scrollTop;
+      const max = container.scrollHeight - container.clientHeight;
+      // Only when the thread really scrolls: a thread shorter than the window
+      // already shows the post at the top, extra padding would just leave a
+      // blank tail.
+      if (max > 0 && target > max) {
+        container.style.paddingBottom = `${Math.ceil(base + (target - max))}px`;
+      } else {
+        container.style.paddingBottom = '';
+      }
       try {
         el.scrollIntoView({ behavior, block: 'start' });
       } catch {
-        const cRect = container.getBoundingClientRect();
-        const eRect = el.getBoundingClientRect();
-        container.scrollTop += eRect.top - cRect.top;
+        container.scrollTop = target;
+      }
+      // Correction pass: `scrollIntoView` can land a few pixels off (scroll
+      // anchoring, sub-pixel padding, momentum scroll). Trust the measured
+      // geometry, not the request.
+      if (behavior === 'auto') {
+        const delta = el.getBoundingClientRect().top - container.getBoundingClientRect().top;
+        if (Math.abs(delta) > 1) container.scrollTop += delta;
       }
     };
     if (typeof requestAnimationFrame === 'function') requestAnimationFrame(apply);
diff --git a/frontend/style.css b/frontend/style.css
index 4829bb7..1f2ec6f 100644
--- a/frontend/style.css
+++ b/frontend/style.css
@@ -9436,10 +9436,12 @@ body.popup-mode .content-area {
 @keyframes bookslm-spin { to { transform: rotate(360deg); } }
 /* Notion-style AI thread: only this zone scrolls; entries stack vertically
    with generous spacing. The latest user post is anchored to the top of the
-   visible area on send (scrollIntoView block:start + scroll-margin). */
+   visible area on send: the post is anchored flush at the top (block:start,
+   and the bottom padding is grown so a short answer can still roll the post
+   all the way up). */
 .bookslm-messages { flex: 1; overflow-y: auto; padding: 20px 16px; display: flex;
-  flex-direction: column; gap: 24px; }
-.bookslm-msg { display: flex; flex-direction: column; gap: 4px; scroll-margin-top: 14px; }
+  flex-direction: column; gap: 24px; overflow-anchor: none; }
+.bookslm-msg { display: flex; flex-direction: column; gap: 4px; }
 /* The user bubble is capped against the *thread* width, not against a
    shrink-to-fit wrapper: capping both would compound (80% × 85% ≈ 68%) and
    produced very short lines. Wrapper stretches, bubble carries the cap. */
diff --git a/frontend/sw.js b/frontend/sw.js
index 5668006..439f0ae 100644
--- a/frontend/sw.js
+++ b/frontend/sw.js
@@ -11,7 +11,7 @@
  * cache or Cloudflare does NOT clear the Service Worker Cache Storage, which is
  * a separate store. Bumping SW_VERSION invalidates it on every release.
  */
-const SW_VERSION = 'v12';
+const SW_VERSION = 'v18';
 const CODE_CACHE = `obsigate-code-${SW_VERSION}`;
 const RUNTIME_CACHE = `obsigate-runtime-${SW_VERSION}`;
 const API_CACHE = `obsigate-api-${SW_VERSION}`;
diff --git a/tests/frontend/ai.test.mjs b/tests/frontend/ai.test.mjs
index f853961..c27540c 100644
--- a/tests/frontend/ai.test.mjs
+++ b/tests/frontend/ai.test.mjs
@@ -235,6 +235,45 @@ async function main() {
       panel.remove();
     });
 
+    await test("a short answer still rolls the post up: bottom padding is grown", () => {
+      // jsdom has no layout, so the geometry is stubbed: the post sits 500px
+      // down the thread while only 300px of scroll room exist. Without the
+      // padding trick the browser clamps and the post stops mid-window.
+      const b = new BooksLM();
+      b._messages = [
+        { role: "user", content: "question" },
+        { role: "assistant", content: "réponse courte" },
+      ];
+      const panel = document.createElement("div");
+      panel.innerHTML = `<div class="bookslm-messages"></div>`;
+      document.body.appendChild(panel);
+      b._panel = panel;
+      const raf = globalThis.requestAnimationFrame;
+      globalThis.requestAnimationFrame = (fn) => fn();
+      try {
+        // Render the messages, then fake the geometry and re-anchor.
+        b._renderMessages();
+        const c = panel.querySelector(".bookslm-messages");
+        const post = panel.querySelector(".bookslm-msg.user");
+        Object.defineProperty(c, "clientHeight", { value: 400, configurable: true });
+        Object.defineProperty(c, "scrollHeight", { value: 700, configurable: true, writable: true });
+        c.getBoundingClientRect = () => ({ top: 0, left: 0, right: 0, bottom: 400, width: 400, height: 400 });
+        post.getBoundingClientRect = () => ({ top: 500, left: 0, right: 0, bottom: 560, width: 300, height: 60 });
+        b._anchorLatestUser({ instant: true });
+        // target (500) - max (300) = 200px of missing room.
+        assert.equal(c.style.paddingBottom, "200px", "padding grown by the missing room");
+
+        // Plenty of room below the post → no padding at all.
+        Object.defineProperty(c, "scrollHeight", { value: 2000, configurable: true, writable: true });
+        c.scrollTop = 0;
+        b._anchorLatestUser({ instant: true });
+        assert.equal(c.style.paddingBottom, "", "no padding when the thread already scrolls far enough");
+      } finally {
+        globalThis.requestAnimationFrame = raf;
+        panel.remove();
+      }
+    });
+
     await test("the placeholder render keeps the anchor (no scroll reset before first token)", async () => {
       // Regression: the assistant placeholder used to be rendered while
       // `_isLoading` was still false, so it took the "restore previous