feat(model-server): progression fine du téléchargement du modèle llamacpp (#54)

Stretch B2/F2 de #54, par-dessus le MVP déjà mergé (B1/F1).

Backend : le port de téléchargement HF publie une progression débouncée
(bytes reçus / total, pourcentage) via le stream de statut du serveur
modèle, avec gestion du total inconnu (pas de faux %), du cache hit,
de l'annulation et du timeout.

Frontend : l'overlay plein-cellule de préparation du serveur affiche la
progression réelle (barre, %, octets, source) en mappant le fil de
statut, avec la règle « pas de faux % » quand le total est inconnu.

Tests : application + infrastructure (téléchargement débouncé, cancel,
timeout, cache hit, total inconnu) et vitest (overlay + formatage pur).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

frontend/src/features/agents/modelServerLaunch.ts

@@ -37,9 +37,9 @@ export function correlateModelServerStatus(
  * - `downloading` → "Téléchargement du modèle…",
  * - `starting`/`probing` → "Chargement du serveur…".
  *
- * F1 MVP ignores the progress fields (bytes/percent) — the bar/% is the F2
- * stretch. This is the single predicate that decides whether the launch overlay
- * covers a cell.
+ * This is the single predicate that decides whether the launch overlay covers a
+ * cell (and supplies its title). The download *progress* (bar/%/bytes, F2) is a
+ * separate, additive concern — see {@link describeModelServerDownload}.
  */
 export function modelServerOverlayText(
   status: ModelServerStatus | undefined,
@@ -56,6 +56,60 @@ export function modelServerOverlayText(
   }
 }
 
+/**
+ * Human-readable byte size, decimal (SI) units so it matches what download UIs
+ * and llama.cpp report (`1 Mo = 1000 Ko`). Bytes are shown whole; larger units
+ * keep one decimal, trailing `.0` trimmed. Guards against non-finite/negative
+ * inputs (→ "0 o").
+ */
+export function formatBytes(n: number): string {
+  if (!Number.isFinite(n) || n <= 0) return "0 o";
+  const units = ["o", "Ko", "Mo", "Go", "To"];
+  let value = n;
+  let unit = 0;
+  while (value >= 1000 && unit < units.length - 1) {
+    value /= 1000;
+    unit += 1;
+  }
+  const rounded =
+    unit === 0 ? Math.round(value) : Math.round(value * 10) / 10;
+  return `${rounded} ${units[unit]}`;
+}
+
+/**
+ * Download progress detail extracted from a `downloading` status (F2), or `null`
+ * for any other state (`starting`/`probing`/`ready`/`failed` → no bar). Mirrors
+ * the wire shape without inventing data:
+ * - `percent`: the backend's completion % clamped to `0..100` when the total is
+ *   known, else `null` → the overlay shows an **indeterminate** bar (no fake %);
+ * - `bytesLabel`: `"X Mo / Y Mo"` when both bounds are known, `"X Mo"` when only
+ *   the downloaded amount is, else `null`;
+ * - `source`: the remote model ref (e.g. HF `namespace/repo`) when present.
+ */
+export interface ModelServerDownloadProgress {
+  percent: number | null;
+  bytesLabel: string | null;
+  source: string | null;
+}
+
+export function describeModelServerDownload(
+  status: ModelServerStatus | undefined,
+): ModelServerDownloadProgress | null {
+  if (!status || status.state !== "downloading") return null;
+  const { downloadedBytes, totalBytes, percent, source } = status;
+  const bytesLabel =
+    downloadedBytes != null && totalBytes != null
+      ? `${formatBytes(downloadedBytes)} / ${formatBytes(totalBytes)}`
+      : downloadedBytes != null
+        ? formatBytes(downloadedBytes)
+        : null;
+  return {
+    percent: percent != null ? Math.min(100, Math.max(0, percent)) : null,
+    bytesLabel,
+    source: source ?? null,
+  };
+}
+
 /** What {@link useModelServerLaunchState} exposes to a consumer. */
 export interface ModelServerLaunchState {
   /**