feat: enhance GPU support detection and user feedback in the engine interface

web/src/components/Shell.tsx

@@ -102,7 +102,22 @@ export function Shell({ children, activeRun }: { children: ReactNode; activeRun:
       <div className="mt-1 truncate text-[0.75rem] font-medium text-ink-200" title={system.engine_runtime}>
         {system.engine_runtime}
       </div>
-      <div className="text-[0.6875rem] text-ink-500">{system.engine_architecture}</div>
+      <div className="text-[0.6875rem] text-ink-500">
+        {system.engine_architecture}
+        {/* The architecture alone is misleading when the installed
+            llama-cpp-python is a CPU build: it reports the hardware, not what
+            the binary can drive. Saying "CPU build" here is the difference
+            between a user seeing "cuda" and assuming the GPU is busy, and
+            knowing their card is idle. */}
+        {system.engine_gpu_offload === false && (
+          <span className="text-amber-400"> · CPU build</span>
+        )}
+      </div>
+      {system.engine_warning && (
+        <div className="mt-1.5 text-[0.6875rem] leading-snug text-amber-400">
+          {system.engine_warning}
+        </div>
+      )}
       {!system.template_ok && (
         <div className="mt-1.5 text-[0.6875rem] text-rose-400">Report template missing</div>
       )}

web/src/lib/api.ts

@@ -214,6 +214,12 @@ export interface SystemInfo {
   version: string
   engine_architecture: string
   engine_runtime: string
+  /** Whether the installed llama-cpp-python was built with GPU offload.
+   *  null when llama-cpp-python is missing or too old to report it — which is
+   *  distinct from a confirmed false. */
+  engine_gpu_offload: boolean | null
+  /** Set when detected hardware and the installed build disagree. */
+  engine_warning: string | null
   template_ok: boolean
   python_version: string
   metrics: Record<string, string>

web/src/pages/SettingsPage.tsx

@@ -213,6 +213,21 @@ export function SettingsPage() {
                   <dl className="space-y-2.5 text-[0.8125rem]">
                     <InfoRow label="Runtime" value={system.data.engine_runtime} mono />
                     <InfoRow label="Architecture" value={system.data.engine_architecture} />
+                    {/* Architecture is detected from hardware; this is what the
+                        installed binary can actually drive. The two disagreeing
+                        is the difference between a GPU working and sitting idle
+                        while the interface still says "cuda". */}
+                    {system.data.engine_gpu_offload !== null && (
+                      <InfoRow
+                        label="GPU offload"
+                        value={
+                          system.data.engine_gpu_offload
+                            ? 'Supported by this build'
+                            : 'Not in this build (CPU only)'
+                        }
+                        tone={system.data.engine_gpu_offload ? 'ok' : 'bad'}
+                      />
+                    )}
                     <InfoRow label="Platform" value={system.data.metrics.platform ?? '—'} />
                     <InfoRow label="Python" value={system.data.python_version} mono />
                     <InfoRow label="LM-Gambit" value={`v${system.data.version}`} mono />