fix(whisper): handle model warmup events

- Ignore backend model lifecycle webhooks so model warmup does not
  mark jobs done early
- Parse batched SSE messages and relay model load states during
  submit retries

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

src/lib/server/whisper.ts

@@ -1,4 +1,32 @@
-import type { ModelStatus } from '$lib/types.js';
+import type { ModelStateTag, ModelStatus } from '$lib/types.js';
+
+const MODEL_STATES = new Set<ModelStateTag>(['unloaded', 'loading', 'waiting_for_gpu', 'ready']);
+
+function isModelStateTag(value: unknown): value is ModelStateTag {
+	return typeof value === 'string' && MODEL_STATES.has(value as ModelStateTag);
+}
+
+function extractSseMessages(buffer: string): { messages: { eventType: string; data: string }[]; rest: string } {
+	const normalized = buffer.replace(/\r/g, '');
+	const chunks = normalized.split('\n\n');
+	const rest = chunks.pop() ?? '';
+	const messages = chunks
+		.map((chunk) => {
+			let eventType = '';
+			const dataLines: string[] = [];
+			for (const line of chunk.split('\n')) {
+				if (line.startsWith('event:')) {
+					eventType = line.slice(6).trim();
+				} else if (line.startsWith('data:')) {
+					dataLines.push(line.slice(5).trim());
+				}
+			}
+			return { eventType, data: dataLines.join('\n') };
+		})
+		.filter((message) => message.data.length > 0);
+
+	return { messages, rest };
+}
 
 function whisperUrl() {
 	return process.env.WHISPER_URL ?? 'http://localhost:8080';
@@ -22,7 +50,10 @@ export async function getModelStatus(): Promise<ModelStatus> {
  * SSE connection fails or closes without that event, so the retry loop can
  * try again without hanging indefinitely.
  */
-async function waitForModelReady(timeoutMs: number): Promise<void> {
+async function waitForModelReady(
+	timeoutMs: number,
+	onStateChange?: (state: ModelStateTag) => void
+): Promise<void> {
 	const { default: fetch } = await import('node-fetch');
 	const ac = new AbortController();
 	return new Promise((resolve) => {
@@ -47,17 +78,18 @@ async function waitForModelReady(timeoutMs: number): Promise<void> {
 				for await (const chunk of res.body) {
 					if (ac.signal.aborted) break;
 					buf += chunk.toString();
-					const lines = buf.split('\n');
-					buf = lines.pop() ?? '';
-					for (const line of lines) {
-						if (!line.startsWith('data:')) continue;
+					const { messages, rest } = extractSseMessages(buf);
+					buf = rest;
+					for (const message of messages) {
 						try {
-							const payload = JSON.parse(line.slice(5).trim());
+							const payload = JSON.parse(message.data) as { state?: unknown };
+							if (!isModelStateTag(payload.state)) continue;
 							if (payload.state === 'ready') {
 								clearTimeout(timer);
 								finish();
 								return;
 							}
+							onStateChange?.(payload.state);
 						} catch { /* ignore parse errors */ }
 					}
 				}
@@ -83,7 +115,7 @@ export async function submitJob(
 	wavPath: string,
 	webhookUrl: string,
 	language?: string,
-	onModelWaiting?: (state: string, retryAfterSecs: number) => void,
+	onModelWaiting?: (state: ModelStateTag, retryAfterSecs: number) => void,
 	maxAttempts = 20
 ): Promise<string> {
 	const FormData = (await import('form-data')).default;
@@ -116,10 +148,15 @@ export async function submitJob(
 				state?: string;
 				retry_after_secs?: number;
 			};
-			const state = body.state ?? 'unloaded';
+			const state = isModelStateTag(body.state) ? body.state : 'unloaded';
 			const waitSecs = body.retry_after_secs ?? parseInt(res.headers.get('Retry-After') ?? '15');
 			onModelWaiting?.(state, waitSecs);
-			await waitForModelReady((waitSecs + 1) * 1000);
+			let lastState = state;
+			await waitForModelReady((waitSecs + 1) * 1000, (nextState) => {
+				if (nextState === lastState) return;
+				lastState = nextState;
+				onModelWaiting?.(nextState, waitSecs);
+			});
 			continue;
 		}
 
@@ -169,30 +206,23 @@ export async function streamJob(
 	let buf = '';
 	for await (const chunk of res.body) {
 		buf += chunk.toString();
-		const lines = buf.split('\n');
-		buf = lines.pop() ?? '';
+		const { messages, rest } = extractSseMessages(buf);
+		buf = rest;
 
-		let eventType = '';
-		let dataLine = '';
-		for (const line of lines) {
-			if (line.startsWith('event:')) eventType = line.slice(6).trim();
-			else if (line.startsWith('data:')) dataLine = line.slice(5).trim();
+		for (const message of messages) {
+			try {
+				const payload = JSON.parse(message.data);
+				if (payload.type === 'progress') {
+					onProgress(payload.percent ?? 0, payload.chunk ?? 0, payload.total ?? 0);
+				} else if (payload.type === 'done') {
+					onDone();
+					return;
+				} else if (payload.type === 'error') {
+					onError(payload.message ?? 'unknown error');
+					return;
+				}
+			} catch { /* ignore parse errors */ }
 		}
-
-		if (!dataLine) continue;
-
-		try {
-			const payload = JSON.parse(dataLine);
-			if (payload.type === 'progress') {
-				onProgress(payload.percent ?? 0, payload.chunk ?? 0, payload.total ?? 0);
-			} else if (payload.type === 'done') {
-				onDone();
-				return;
-			} else if (payload.type === 'error') {
-				onError(payload.message ?? 'unknown error');
-				return;
-			}
-		} catch { /* ignore parse errors */ }
 	}
 }