Add relaxed-MTP llama-server as a new connection, surface generation speed

Wires the custom relaxed-acceptance llama-server (see
../mtp-relaxed-decoding) into the stack as a new llama-mtp service,
registered as an additional OpenAI-compatible connection alongside
Ollama. ollama-auth now proxies to llama-mtp instead of the now-empty
Ollama, so LAN clients (e.g. Home Assistant's voice pipeline) keep
working against the same URL/token with no reconfiguration.

Also merges llama.cpp's `timings` extension (dropped by the generic
OpenAI response schema) into message.usage and surfaces it as a
labeled tok/s badge, so relaxed-MTP responses get the same visible
generation-speed info Ollama responses already had.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

src/lib/components/chat/Chat.svelte

@@ -2744,7 +2744,8 @@
 	};
 
 	const chatCompletionEventHandler = async (data, message, chatId) => {
-		const { id, done, choices, content, output, sources, selected_model_id, error, usage } = data;
+		const { id, done, choices, content, output, sources, selected_model_id, error, usage, timings } =
+			data;
 
 		// Store raw OR-aligned output items from backend
 		if (output) {
@@ -2797,8 +2798,12 @@
 			message.arena = true;
 		}
 
-		if (usage) {
-			message.usage = usage;
+		if (usage || timings) {
+			// llama.cpp-compatible servers (e.g. our relaxed-MTP llama-server) send
+			// generation speed as a sibling `timings` object rather than inside
+			// `usage` -- merge it in so it surfaces in the same response-info
+			// tooltip Ollama's eval_count/eval_duration fields already populate.
+			message.usage = { ...(usage || {}), ...(timings || {}) };
 		}
 
 		history.messages[message.id] = message;

src/lib/components/chat/Messages/ResponseMessage.svelte

@@ -144,6 +144,23 @@
 		}
 	}
 
+	// Generation speed, normalized across backends: Ollama reports
+	// eval_count/eval_duration (nanoseconds), llama.cpp-compatible servers
+	// (e.g. our relaxed-MTP llama-server) report timings.predicted_per_second
+	// directly -- note that field name is llama.cpp's generic term for "the
+	// main generation loop's output," i.e. the actual response stream, NOT
+	// the MTP drafter in isolation (which is separately broken out as
+	// draft_n/draft_n_accepted).
+	const getTokensPerSecond = (usage: Record<string, unknown> | undefined | null) => {
+		if (!usage) return null;
+		if (typeof usage.predicted_per_second === 'number') return usage.predicted_per_second;
+		if (typeof usage.eval_count === 'number' && typeof usage.eval_duration === 'number' && usage.eval_duration > 0) {
+			return (usage.eval_count / usage.eval_duration) * 1e9;
+		}
+		return null;
+	};
+	$: tokensPerSecond = getTokensPerSecond(message.usage);
+
 	export let siblings;
 
 	export let setInputText: Function = () => {};
@@ -1189,6 +1206,14 @@
 									</Tooltip>
 								{/if}
 
+								{#if tokensPerSecond}
+									<span
+										class="text-xs text-gray-400 dark:text-gray-500 px-1 self-center whitespace-nowrap"
+									>
+										{tokensPerSecond.toFixed(1)} tok/s
+									</span>
+								{/if}
+
 								{#if message.usage}
 									<Tooltip
 										content={message.usage