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>

.gitea/workflows/docker-build.yml

@@ -15,8 +15,28 @@ env:
   IMAGE_NAME: mozempk/tonemark
 
 jobs:
+  test:
+    runs-on: ubuntu-latest
+
+    steps:
+      - name: Checkout repository
+        uses: actions/checkout@v4
+
+      - name: Set up Node.js
+        uses: actions/setup-node@v4
+        with:
+          node-version: '22'
+          cache: 'npm'
+
+      - name: Install dependencies
+        run: npm install
+
+      - name: Run tests
+        run: npm test
+
   build-and-push:
     runs-on: ubuntu-latest
+    needs: test
 
     steps:
       - name: Checkout repository

package-lock.json

@@ -12,7 +12,8 @@
         "better-sqlite3": "^12.9.0",
         "form-data": "^4.0.5",
         "node-fetch": "^3.3.2",
-        "web-push": "^3.6.7"
+        "web-push": "^3.6.7",
+        "youtube-transcript": "^1.3.1"
       },
       "devDependencies": {
         "@sveltejs/adapter-auto": "^7.0.1",
@@ -89,6 +90,27 @@
         "node": ">=18"
       }
     },
+    "node_modules/@emnapi/core": {
+      "version": "1.10.0",
+      "resolved": "https://registry.npmjs.org/@emnapi/core/-/core-1.10.0.tgz",
+      "integrity": "sha512-yq6OkJ4p82CAfPl0u9mQebQHKPJkY7WrIuk205cTYnYe+k2Z8YBh11FrbRG/H6ihirqcacOgl2BIO8oyMQLeXw==",
+      "license": "MIT",
+      "optional": true,
+      "dependencies": {
+        "@emnapi/wasi-threads": "1.2.1",
+        "tslib": "^2.4.0"
+      }
+    },
+    "node_modules/@emnapi/runtime": {
+      "version": "1.10.0",
+      "resolved": "https://registry.npmjs.org/@emnapi/runtime/-/runtime-1.10.0.tgz",
+      "integrity": "sha512-ewvYlk86xUoGI0zQRNq/mC+16R1QeDlKQy21Ki3oSYXNgLb45GV1P6A0M+/s6nyCuNDqe5VpaY84BzXGwVbwFA==",
+      "license": "MIT",
+      "optional": true,
+      "dependencies": {
+        "tslib": "^2.4.0"
+      }
+    },
     "node_modules/@emnapi/wasi-threads": {
       "version": "1.2.1",
       "resolved": "https://registry.npmjs.org/@emnapi/wasi-threads/-/wasi-threads-1.2.1.tgz",
@@ -896,7 +918,6 @@
       "resolved": "https://registry.npmjs.org/@sveltejs/kit/-/kit-2.59.1.tgz",
       "integrity": "sha512-d8OON70AphLdDesuTIl//M2O6fRTIicX8aYv8vhCiYEhTTI2OboKqey0Hu1A4VFhqwgqtq0vKDmPFGkw8kKmgw==",
       "license": "MIT",
-      "peer": true,
       "dependencies": {
         "@standard-schema/spec": "^1.0.0",
         "@sveltejs/acorn-typescript": "^1.0.5",
@@ -938,7 +959,6 @@
       "resolved": "https://registry.npmjs.org/@sveltejs/vite-plugin-svelte/-/vite-plugin-svelte-7.1.1.tgz",
       "integrity": "sha512-FOJdbE5pxae68DoTBJ49t1dIA7TSmMHR6CsuJhX90cO/UfrEMHA7KJNUj3WdZuUDJPu4ujqpJ2Tgqd2gTWr6Xg==",
       "license": "MIT",
-      "peer": true,
       "dependencies": {
         "deepmerge": "^4.3.1",
         "magic-string": "^0.30.21",
@@ -1313,7 +1333,6 @@
       "integrity": "sha512-38C0/Ddb7HcRG0Z4/DUem8x57d2p9jYgp18mkaYswEOQBGsI1CG4f/hjm0ZCeaJfWhSZ4k7jgs29V1Zom7Ki9A==",
       "dev": true,
       "license": "MIT",
-      "peer": true,
       "dependencies": {
         "@bcoe/v8-coverage": "^1.0.2",
         "@vitest/utils": "4.1.5",
@@ -1467,7 +1486,6 @@
       "resolved": "https://registry.npmjs.org/acorn/-/acorn-8.16.0.tgz",
       "integrity": "sha512-UVJyE9MttOsBQIDKw1skb9nAwQuR5wuGD3+82K6JgJlm/Y+KI92oNsMNGZCYdDsVtRHSak0pcV5Dno5+4jh9sw==",
       "license": "MIT",
-      "peer": true,
       "bin": {
         "acorn": "bin/acorn"
       },
@@ -3021,7 +3039,6 @@
       "resolved": "https://registry.npmjs.org/rollup/-/rollup-4.60.3.tgz",
       "integrity": "sha512-pAQK9HalE84QSm4Po3EmWIZPd3FnjkShVkiMlz1iligWYkWQ7wHYd1PF/T7QZ5TVSD6uSTon5gBVMSM4JfBV+A==",
       "license": "MIT",
-      "peer": true,
       "dependencies": {
         "@types/estree": "1.0.8"
       },
@@ -3255,7 +3272,6 @@
       "resolved": "https://registry.npmjs.org/svelte/-/svelte-5.55.5.tgz",
       "integrity": "sha512-2uCs/LZ9us+AktdzYJM8OcxQ8qnPS1kpaO7syGT/MgO+6Qr1Ybl+TqPq+97u7PHqmmMlye5ZkoyXONy5mjjAbw==",
       "license": "MIT",
-      "peer": true,
       "dependencies": {
         "@jridgewell/remapping": "^2.3.4",
         "@jridgewell/sourcemap-codec": "^1.5.0",
@@ -3428,7 +3444,6 @@
       "integrity": "sha512-y2TvuxSZPDyQakkFRPZHKFm+KKVqIisdg9/CZwm9ftvKXLP8NRWj38/ODjNbr43SsoXqNuAisEf1GdCxqWcdBw==",
       "devOptional": true,
       "license": "Apache-2.0",
-      "peer": true,
       "bin": {
         "tsc": "bin/tsc",
         "tsserver": "bin/tsserver"
@@ -3455,7 +3470,6 @@
       "resolved": "https://registry.npmjs.org/vite/-/vite-8.0.10.tgz",
       "integrity": "sha512-rZuUu9j6J5uotLDs+cAA4O5H4K1SfPliUlQwqa6YEwSrWDZzP4rhm00oJR5snMewjxF5V/K3D4kctsUTsIU9Mw==",
       "license": "MIT",
-      "peer": true,
       "dependencies": {
         "lightningcss": "^1.32.0",
         "picomatch": "^4.0.4",
@@ -3553,7 +3567,6 @@
       "integrity": "sha512-9Xx1v3/ih3m9hN+SbfkUyy0JAs72ap3r7joc87XL6jwF0jGg6mFBvQ1SrwaX+h8BlkX6Hz9shdd1uo6AF+ZGpg==",
       "dev": true,
       "license": "MIT",
-      "peer": true,
       "dependencies": {
         "@vitest/expect": "4.1.5",
         "@vitest/mocker": "4.1.5",
@@ -3689,6 +3702,15 @@
       "integrity": "sha512-l4Sp/DRseor9wL6EvV2+TuQn63dMkPjZ/sp9XkghTEbV9KlPS1xUsZ3u7/IQO4wxtcFB4bgpQPRcR3QCvezPcQ==",
       "license": "ISC"
     },
+    "node_modules/youtube-transcript": {
+      "version": "1.3.1",
+      "resolved": "https://registry.npmjs.org/youtube-transcript/-/youtube-transcript-1.3.1.tgz",
+      "integrity": "sha512-NDCjwad113TGybbYF51y9Z4tcwzBHUZWQdF9veULNca18L+FdDbHHtTHIr69WVa3bB90l67S8kN0HtL2JO9fhg==",
+      "license": "MIT",
+      "engines": {
+        "node": ">=18.0.0"
+      }
+    },
     "node_modules/zimmerframe": {
       "version": "1.1.4",
       "resolved": "https://registry.npmjs.org/zimmerframe/-/zimmerframe-1.1.4.tgz",

package.json

@@ -34,6 +34,7 @@
     "better-sqlite3": "^12.9.0",
     "form-data": "^4.0.5",
     "node-fetch": "^3.3.2",
-    "web-push": "^3.6.7"
+    "web-push": "^3.6.7",
+    "youtube-transcript": "^1.3.1"
   }
 }

src/lib/components/RecordButton.svelte

@@ -16,10 +16,10 @@
 
 	type RecordState = 'idle' | 'requesting' | 'recording' | 'stopping';
 
-	let state = $state<RecordState>('idle');
-	let error = $state('');
-	let elapsed = $state(0); // seconds
-	let liveData = $state<Float32Array | null>(null);
+	let recordState: RecordState = $state('idle');
+	let error: string = $state('');
+	let elapsed: number = $state(0); // seconds
+	let liveData: Float32Array | null = $state(null);
 
 	let mediaRecorder: MediaRecorder | null = null;
 	let chunks: Blob[] = [];
@@ -60,12 +60,12 @@
 
 	async function startRecording() {
 		error = '';
-		state = 'requesting';
+		recordState = 'requesting';
 		try {
 			stream = await navigator.mediaDevices.getUserMedia({ audio: true });
 		} catch {
 			error = 'Microphone access denied';
-			state = 'idle';
+			recordState = 'idle';
 			return;
 		}
 
@@ -81,11 +81,11 @@
 
 		elapsed = 0;
 		timerInterval = setInterval(() => elapsed++, 1000);
-		state = 'recording';
+		recordState = 'recording';
 	}
 
 	function stopRecording() {
-		state = 'stopping';
+		recordState = 'stopping';
 		mediaRecorder?.stop();
 		if (timerInterval) clearInterval(timerInterval);
 		if (animFrame) cancelAnimationFrame(animFrame);
@@ -99,7 +99,7 @@
 		const ext = mime.includes('ogg') ? 'ogg' : mime.includes('mp4') ? 'mp4' : 'webm';
 		const blob = new Blob(chunks, { type: mime });
 		const filename = `recording-${new Date().toISOString().slice(0, 19).replace(/[T:]/g, '-')}.${ext}`;
-		state = 'idle';
+		recordState = 'idle';
 		ondone?.(blob, filename);
 	}
 
@@ -116,15 +116,18 @@
 		{ length: IDLE_BARS },
 		(_, i) => 3 + Math.abs(Math.sin(i * 0.7) + Math.cos(i * 0.31)) * 20
 	);
+	const liveBars = $derived.by<number[]>(() =>
+		liveData ? Array.from(liveData.slice(0, IDLE_BARS), (value) => Number(value)) : []
+	);
 </script>
 
 <div class="recorder">
 	<!-- Waveform display -->
 	<div class="waveform-area" aria-hidden="true">
-		{#if state === 'recording' && liveData}
+		{#if recordState === 'recording' && liveData}
 			<!-- Live waveform from AnalyserNode -->
 			<svg viewBox="0 0 {IDLE_BARS * 5} 28" preserveAspectRatio="none" class="waveform-svg">
-				{#each Array.from(liveData).slice(0, IDLE_BARS) as v, i}
+				{#each liveBars as v, i}
 					{@const h = 2 + v * 24}
 					<rect
 						x={i * 5}
@@ -147,8 +150,8 @@
 						width="3"
 						height={h}
 						rx="1.5"
-						fill={state === 'idle' ? 'rgba(255,255,255,0.15)' : accent}
-						opacity={state === 'idle' ? 1 : 0.3}
+						fill={recordState === 'idle' ? 'rgba(255,255,255,0.15)' : accent}
+						opacity={recordState === 'idle' ? 1 : 0.3}
 					/>
 				{/each}
 			</svg>
@@ -156,7 +159,7 @@
 	</div>
 
 	<!-- Timer (recording only) -->
-	{#if state === 'recording'}
+	{#if recordState === 'recording'}
 		<div class="timer" style="color: {accent}">
 			<span class="rec-dot" style="background: {accent}"></span>
 			{formatTime(elapsed)}
@@ -170,15 +173,15 @@
 
 	<!-- Buttons -->
 	<div class="btn-row">
-		{#if state === 'idle' || state === 'requesting'}
+		{#if recordState === 'idle' || recordState === 'requesting'}
 			<button
 				class="btn-record"
 				style="background: {accent}; color: #0c0d10;"
 				onclick={startRecording}
-				disabled={state === 'requesting'}
+				disabled={recordState === 'requesting'}
 				aria-label="Start recording"
 			>
-				{#if state === 'requesting'}
+				{#if recordState === 'requesting'}
 					<svg width="13" height="13" viewBox="0 0 13 13" style="animation: spin 1s linear infinite">
 						<circle cx="6.5" cy="6.5" r="5" stroke="currentColor" stroke-width="1.5" fill="none" stroke-dasharray="20 12"/>
 					</svg>
@@ -190,7 +193,7 @@
 					Record
 				{/if}
 			</button>
-		{:else if state === 'recording'}
+		{:else if recordState === 'recording'}
 			<button
 				class="btn-stop"
 				onclick={stopRecording}

src/lib/job-progress.ts

@@ -0,0 +1,52 @@
+import type { Job, JobStatus } from '$lib/types.js';
+
+export const TERMINAL_JOB_STATUSES: readonly JobStatus[] = ['done', 'failed', 'cancelled'];
+
+const STATUS_LABELS: Record<JobStatus, string> = {
+	pending: 'Pending',
+	downloading: 'Downloading',
+	preparing: 'Preparing',
+	warming_model: 'Loading model',
+	transcribing: 'Transcribing',
+	processing: 'Processing',
+	done: 'Done',
+	failed: 'Failed',
+	cancelled: 'Cancelled'
+};
+
+const STATUS_COLORS: Record<JobStatus, string> = {
+	done: '#cdf24e',
+	failed: '#ff6b6b',
+	cancelled: 'rgba(232,233,236,0.3)',
+	processing: '#76daa2',
+	transcribing: '#80c7f7',
+	warming_model: '#76daa2',
+	preparing: '#fbc94b',
+	downloading: '#a78bfa',
+	pending: 'rgba(232,233,236,0.4)'
+};
+
+export function isTerminalJobStatus(status: JobStatus): boolean {
+	return TERMINAL_JOB_STATUSES.includes(status);
+}
+
+export function getJobStatusLabel(status: JobStatus): string {
+	return STATUS_LABELS[status];
+}
+
+export function getJobStatusColor(status: JobStatus): string {
+	return STATUS_COLORS[status];
+}
+
+export function getDisplayJobProgress(
+	job: Pick<Job, 'status' | 'progress' | 'segmentsJson'>,
+	options: { hasTranscript?: boolean } = {}
+): number {
+	const progress = Math.max(0, Math.min(100, Math.round(job.progress)));
+
+	if (job.status === 'warming_model') return Math.min(progress, 15);
+	if (!isTerminalJobStatus(job.status)) return Math.min(progress, 99);
+	if (job.status === 'done' && !options.hasTranscript) return Math.min(progress, 99);
+
+	return progress;
+}

src/lib/server/downloader.ts

@@ -1,8 +1,9 @@
 import { execFile } from 'child_process';
 import { promisify } from 'util';
 import { existsSync } from 'fs';
-import { mkdir, unlink, writeFile } from 'fs/promises';
+import { mkdir, writeFile } from 'fs/promises';
 import { join } from 'path';
+import { fetchTranscript, type TranscriptResponse } from 'youtube-transcript';
 
 const execFileAsync = promisify(execFile);
 const TMP_DIR = join(process.env.DATA_DIR ?? '/tmp/.whisper-pwa', 'downloads');
@@ -26,43 +27,33 @@ export interface AudioResult {
 export type DownloadResult = CaptionResult | AudioResult;
 
 /** Try to get auto-generated captions from YouTube. Returns null if unavailable. */
-async function tryGetCaptions(url: string, outDir: string): Promise<CaptionResult | null> {
-	const jsonPath = join(outDir, 'info.json');
+async function tryGetCaptions(url: string, _outDir: string): Promise<CaptionResult | null> {
 	try {
-		await execFileAsync('yt-dlp', [
-			'--write-auto-subs',
-			'--sub-langs', 'en.*',
-			'--skip-download',
-			'--write-info-json',
-			'--no-playlist',
-			'-o', join(outDir, '%(title)s.%(ext)s'),
-			url
-		]);
-
-		// Find the VTT/SRT file
-		const { readdirSync } = await import('fs');
-		const files = readdirSync(outDir);
-		const vttFile = files.find((f) => f.endsWith('.vtt') || f.endsWith('.srt'));
-		if (!vttFile) return null;
-
-		let title = 'Untitled';
-		if (existsSync(jsonPath)) {
-			try {
-				const info = JSON.parse((await import('fs')).readFileSync(jsonPath, 'utf8'));
-				title = info.title ?? title;
-			} catch { /* ignore */ }
-		}
-
-		const content = (await import('fs')).readFileSync(join(outDir, vttFile), 'utf8');
-		const segments = parseVtt(content);
+		const transcript = await fetchTranscript(url, { lang: 'en' });
+		const segments = transcriptEntriesToSegments(transcript);
 		if (segments.length === 0) return null;
 
+		const title = await getYouTubeTitle(url);
 		return { type: 'captions', segments, title };
 	} catch {
 		return null;
 	}
 }
 
+async function getYouTubeTitle(url: string): Promise<string> {
+	try {
+		const { stdout } = await execFileAsync('yt-dlp', [
+			'--dump-single-json',
+			'--skip-download',
+			'--no-playlist',
+			url
+		]);
+		return JSON.parse(stdout).title ?? 'Untitled';
+	} catch {
+		return 'Untitled';
+	}
+}
+
 /** Download best audio from YouTube. Returns path to audio file. */
 async function downloadAudio(url: string, outDir: string): Promise<{ audioPath: string; title: string }> {
 	await execFileAsync('yt-dlp', [
@@ -124,39 +115,22 @@ export async function cleanupJobTmp(jobId: string) {
 	} catch { /* ignore */ }
 }
 
-/** Parse a WebVTT string into segments. */
-function parseVtt(
-	content: string
+export function transcriptEntriesToSegments(
+	entries: TranscriptResponse[]
 ): Array<{ index: number; start: number; end: number; text: string; words: [] }> {
-	const segments: Array<{ index: number; start: number; end: number; text: string; words: [] }> = [];
-	const blocks = content.split(/\n\n+/);
-	let index = 0;
-
-	for (const block of blocks) {
-		const lines = block.trim().split('\n');
-		const timeLine = lines.find((l) => l.includes('-->'));
-		if (!timeLine) continue;
-
-		const [startStr, endStr] = timeLine.split('-->').map((s) => s.trim().split(' ')[0]);
-		const start = vttTimeToSec(startStr);
-		const end = vttTimeToSec(endStr);
-		const text = lines
-			.filter((l) => !l.includes('-->') && !/^\d+$/.test(l.trim()) && l.trim())
-			.join(' ')
-			.replace(/<[^>]+>/g, '')
-			.trim();
-
-		if (text) {
-			segments.push({ index: index++, start, end, text, words: [] });
-		}
-	}
-
-	return segments;
-}
-
-function vttTimeToSec(t: string): number {
-	const parts = t.split(':').map(Number);
-	if (parts.length === 3) return parts[0] * 3600 + parts[1] * 60 + parts[2];
-	if (parts.length === 2) return parts[0] * 60 + parts[1];
-	return parts[0];
+	const useMilliseconds = entries.some((entry) => entry.offset > 1000 || entry.duration > 1000);
+	return entries
+		.map((entry) => {
+			const start = useMilliseconds ? entry.offset / 1000 : entry.offset;
+			const duration = useMilliseconds ? entry.duration / 1000 : entry.duration;
+			return {
+				index: 0,
+				start,
+				end: start + duration,
+				text: entry.text.trim(),
+				words: [] as []
+			};
+		})
+		.filter((entry) => entry.text.length > 0)
+		.map((entry, index) => ({ ...entry, index }));
 }

src/lib/server/pipeline.ts

@@ -96,15 +96,13 @@ async function runJob(
 
 		if (captionSegments) {
 			// Caption fast path — skip whisper
-			const { deduplicateSegments } = await import('./postprocess.js');
 			const { writeOutputs } = await import('./formatter.js');
-			const segments = deduplicateSegments(captionSegments);
-			const paths = await writeOutputs(segments, title, jobId);
+			const paths = await writeOutputs(captionSegments, title, jobId);
 			updateJob({
 				id: jobId,
 				status: 'done',
 				progress: 100,
-				segmentsJson: JSON.stringify(segments),
+				segmentsJson: JSON.stringify(captionSegments),
 				outputDir: paths.srt.replace(/\/[^/]+$/, '')
 			});
 			emitProgress(jobId, { type: 'done' });
@@ -126,13 +124,16 @@ async function runJob(
 
 		// ── 4. Submit to whisper with webhook ────────────────────────────────
 		setJobStatus(jobId, 'transcribing', 10);
-		emitProgress(jobId, { type: 'status', status: 'transcribing' });
+		emitProgress(jobId, { type: 'status', status: 'transcribing', progress: 10 });
 
 		const webhookUrl = `${WEBHOOK_BASE_URL}/api/webhook/${jobId}`;
 		const whisperJobId = await submitJob(wavPath, webhookUrl, language, (state, retryAfterSecs) => {
-			emitProgress(jobId, { type: 'model_warming', state, retryAfterSecs });
+			setJobStatus(jobId, 'warming_model', 10);
+			emitProgress(jobId, { type: 'model_warming', status: 'warming_model', state, retryAfterSecs, progress: 10 });
 		});
 		updateJob({ id: jobId, whisperJobId });
+		setJobStatus(jobId, 'transcribing', 10);
+		emitProgress(jobId, { type: 'status', status: 'transcribing', progress: 10 });
 
 		// ── 5. Open SSE for live progress (non-blocking relay) ───────────────
 		streamJob(

src/lib/server/postprocess.ts

@@ -1,108 +0,0 @@
-import type { Segment } from '$lib/types.js';
-
-// ── Collapse consecutive repeated phrases within a segment's text ────────────
-
-function collapseRepeats(text: string): string {
-	let prev = '';
-	// Keep applying until stable
-	while (true) {
-		const next = collapseOnce(text);
-		if (next === prev || next === text) return next;
-		prev = text;
-		text = next;
-	}
-}
-
-function collapseOnce(text: string): string {
-	// Match any repeated phrase (2+ words) appearing consecutively
-	return text.replace(/\b(.{10,}?)\s+\1\b/gi, '$1');
-}
-
-// ── Merge consecutive segments with identical (or near-identical) text ───────
-
-function normalise(s: string) {
-	return s.toLowerCase().replace(/[^\w\s]/g, '').replace(/\s+/g, ' ').trim();
-}
-
-function mergeConsecutive(segments: Segment[]): Segment[] {
-	const out: Segment[] = [];
-	for (const seg of segments) {
-		const last = out[out.length - 1];
-		if (last && normalise(last.text) === normalise(seg.text)) {
-			last.end = seg.end;
-		} else {
-			out.push({ ...seg });
-		}
-	}
-	return out;
-}
-
-// ── N-gram deduplication ─────────────────────────────────────────────────────
-
-const NGRAM_N = 6;
-const LOOKBACK_CHARS = 500;
-const SIMILARITY_THRESHOLD = 0.6;
-
-function ngrams(text: string, n: number): string[] {
-	const words = text.toLowerCase().split(/\s+/);
-	const grams: string[] = [];
-	for (let i = 0; i <= words.length - n; i++) {
-		grams.push(words.slice(i, i + n).join(' '));
-	}
-	return grams;
-}
-
-function jaccardSimilarity(a: string, b: string): number {
-	const ga = new Set(ngrams(a, NGRAM_N));
-	const gb = new Set(ngrams(b, NGRAM_N));
-	// If neither text is long enough to produce n-grams they cannot be compared;
-	// treat as dissimilar so short segments are never incorrectly discarded.
-	if (ga.size === 0 && gb.size === 0) return 0;
-	const intersection = [...ga].filter((g) => gb.has(g)).length;
-	const union = new Set([...ga, ...gb]).size;
-	return union === 0 ? 0 : intersection / union;
-}
-
-function ngramDedup(segments: Segment[]): Segment[] {
-	const out: Segment[] = [];
-	for (const seg of segments) {
-		const windowText = out
-			.slice(-20)
-			.map((s) => s.text)
-			.join(' ')
-			.slice(-LOOKBACK_CHARS);
-
-		if (windowText.length > 0 && jaccardSimilarity(seg.text, windowText) >= SIMILARITY_THRESHOLD) {
-			continue; // duplicate — skip
-		}
-		out.push(seg);
-	}
-	return out;
-}
-
-// ── Full deduplication pipeline ──────────────────────────────────────────────
-
-export function deduplicateSegments(segments: Segment[]): Segment[] {
-	// 1. Collapse repeats within each segment's text
-	let result = segments.map((s) => ({
-		...s,
-		text: collapseRepeats(s.text.trim())
-	}));
-
-	// 2. Remove empty segments
-	result = result.filter((s) => s.text.length > 0);
-
-	// 3. First merge pass
-	result = mergeConsecutive(result);
-
-	// 4. N-gram dedup
-	result = ngramDedup(result);
-
-	// 5. Second merge pass (catches new adjacencies after dedup)
-	result = mergeConsecutive(result);
-
-	// 6. Re-index
-	result.forEach((s, i) => (s.index = i));
-
-	return result;
-}

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 */ }
 					}
 				}
@@ -83,20 +115,23 @@ 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;
 	const { createReadStream } = await import('fs');
 	const { default: fetch } = await import('node-fetch');
 
+	for (let attempt = 1; attempt <= maxAttempts; attempt++) {
+		// Recreate form with a fresh readable stream on every attempt.
+		// A consumed ReadStream cannot be rewound, so reusing it across retries
+		// would send an empty body to whisper after the first 503.
 		const form = new FormData();
 		form.append('audio', createReadStream(wavPath));
 		form.append('task', 'transcribe');
 		form.append('webhook_url', webhookUrl);
 		if (language) form.append('language', language);
 
-	for (let attempt = 1; attempt <= maxAttempts; attempt++) {
 		const res = await fetch(`${whisperUrl()}/jobs`, {
 			method: 'POST',
 			body: form,
@@ -113,10 +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;
 		}
 
@@ -127,6 +167,17 @@ export async function submitJob(
 	throw new Error(`Whisper model did not become ready after ${maxAttempts} attempts`);
 }
 
+/** Unload the model from VRAM. Throws if the whisper server returns non-ok. */
+export async function unloadModel(): Promise<{ ok: boolean }> {
+	const { default: fetch } = await import('node-fetch');
+	const res = await fetch(`${whisperUrl()}/model/unload`, {
+		method: 'POST',
+		signal: AbortSignal.timeout(10000)
+	});
+	if (!res.ok) throw new Error(`/model/unload returned ${res.status}`);
+	return res.json() as Promise<{ ok: boolean }>;
+}
+
 /**
  * Cancel a queued or running job on the whisper server (best-effort).
  * Errors are silently ignored — local job status is already set to cancelled.
@@ -155,20 +206,12 @@ export async function streamJob(
 	let buf = '';
 	for await (const chunk of res.body) {
 		buf += chunk.toString();
-		const lines = buf.split('\n');
-		buf = lines.pop() ?? '';
-
-		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();
-		}
-
-		if (!dataLine) continue;
+		const { messages, rest } = extractSseMessages(buf);
+		buf = rest;
 
+		for (const message of messages) {
 			try {
-			const payload = JSON.parse(dataLine);
+				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') {
@@ -180,6 +223,7 @@ export async function streamJob(
 				}
 			} catch { /* ignore parse errors */ }
 		}
+	}
 }
 
 /** Check if the whisper server is healthy. */

src/lib/types.ts

@@ -12,7 +12,16 @@ export interface ModelStatus {
 	vram_total_mb?: number;
 }
 
-export type JobStatus = 'pending' | 'downloading' | 'preparing' | 'transcribing' | 'processing' | 'done' | 'failed' | 'cancelled';
+export type JobStatus =
+	| 'pending'
+	| 'downloading'
+	| 'preparing'
+	| 'warming_model'
+	| 'transcribing'
+	| 'processing'
+	| 'done'
+	| 'failed'
+	| 'cancelled';
 
 export interface Segment {
 	index: number;

src/routes/+page.svelte

@@ -1,6 +1,7 @@
 <script lang="ts">
 	import { onMount } from 'svelte';
 	import type { Job, AudioMode } from '$lib/types.js';
+	import { getDisplayJobProgress, getJobStatusLabel } from '$lib/job-progress.js';
 	import SourceIcon from '$lib/components/SourceIcon.svelte';
 	import Waveform from '$lib/components/Waveform.svelte';
 	import RecordButton from '$lib/components/RecordButton.svelte';
@@ -95,8 +96,7 @@
 		const parts: string[] = [];
 		if (job.source && !job.source.startsWith('http')) parts.push(job.source.split('/').pop() ?? '');
 		if (job.audioMode) parts.push(job.audioMode);
-		if (job.status === 'done') parts.push('done');
-		else parts.push(job.status);
+		parts.push(getJobStatusLabel(job.status).toLowerCase());
 		return parts.join(' · ');
 	}
 
@@ -144,7 +144,7 @@
 
 			<!-- Decorative waveform -->
 			<div class="dropzone-wave">
-				<Waveform bars={DROPZONE_BARS} progress={0} {ACCENT} height={38} />
+				<Waveform bars={DROPZONE_BARS} progress={0} accent={ACCENT} height={38} />
 			</div>
 
 			<input
@@ -271,7 +271,9 @@
 							<div class="recent-meta mono">{jobMeta(job)}</div>
 						</div>
 						{#if job.status !== 'done' && job.status !== 'failed' && job.status !== 'cancelled'}
-							<div class="recent-progress mono" style="color: {ACCENT}">{job.progress}%</div>
+							<div class="recent-progress mono" style="color: {ACCENT}">
+								{getDisplayJobProgress(job, { hasTranscript: Boolean(job.segmentsJson) })}%
+							</div>
 						{/if}
 						<svg width="14" height="14" viewBox="0 0 14 14" style="color: var(--text-dim); flex-shrink:0">
 							<path d="M5 3l4 4-4 4" stroke="currentColor" stroke-width="1.4" fill="none" stroke-linecap="round" stroke-linejoin="round"/>
@@ -586,4 +588,3 @@
 		}
 	}
 </style>
-

src/routes/api/jobs/[id]/+server.ts

@@ -9,7 +9,7 @@ export async function GET({ params }) {
 	return json(job);
 }
 
-const ACTIVE = new Set(['pending', 'downloading', 'preparing', 'transcribing', 'processing']);
+const ACTIVE = new Set(['pending', 'downloading', 'preparing', 'warming_model', 'transcribing', 'processing']);
 
 export async function DELETE({ params }) {
 	const job = getJob(params.id);

src/routes/api/jobs/[id]/reprocess/+server.ts

@@ -1,10 +1,9 @@
 import { json, error } from '@sveltejs/kit';
 import { getJob, updateJob } from '$lib/server/db.js';
-import { deduplicateSegments } from '$lib/server/postprocess.js';
 import { writeOutputs } from '$lib/server/formatter.js';
 import type { Segment } from '$lib/types.js';
 
-/** POST /api/jobs/[id]/reprocess — re-run post-processing and regenerate all output files. */
+/** POST /api/jobs/[id]/reprocess — regenerate output files from stored canonical segments. */
 export async function POST({ params }) {
 	const job = getJob(params.id);
 	if (!job) throw error(404, 'Job not found');
@@ -14,8 +13,7 @@ export async function POST({ params }) {
 	}
 
 	try {
-		const rawSegments = JSON.parse(job.segmentsJson) as Segment[];
-		const segments = deduplicateSegments(rawSegments);
+		const segments = JSON.parse(job.segmentsJson) as Segment[];
 
 		const paths = await writeOutputs(segments, job.title, job.id);
 		const outputDir = paths.srt.replace(/\/[^/]+$/, '');

src/routes/api/model/unload/+server.ts

@@ -0,0 +1,13 @@
+import { json } from '@sveltejs/kit';
+import { unloadModel } from '$lib/server/whisper.js';
+
+/** Proxy for POST /model/unload on the whisper backend. */
+export async function POST() {
+	try {
+		const body = await unloadModel();
+		return json(body);
+	} catch (err: unknown) {
+		const message = err instanceof Error ? err.message : String(err);
+		return json({ ok: false, error: message }, { status: 502 });
+	}
+}

src/routes/api/webhook/[jobId]/+server.ts

@@ -1,59 +1,90 @@
 import { json, error } from '@sveltejs/kit';
 import { getJob, updateJob, setJobStatus } from '$lib/server/db.js';
-import { deduplicateSegments } from '$lib/server/postprocess.js';
 import { writeOutputs } from '$lib/server/formatter.js';
 import { sendNotification } from '$lib/server/push.js';
 import { cleanupJobTmp } from '$lib/server/downloader.js';
 import { emitProgress } from '$lib/server/pipeline.js';
 import type { Segment, WhisperJob } from '$lib/types.js';
 
+const WHISPER_JOB_STATUSES = new Set<WhisperJob['status']>([
+	'queued',
+	'running',
+	'done',
+	'failed',
+	'cancelled'
+]);
+
+function isWhisperJobWebhook(payload: unknown): payload is WhisperJob {
+	if (!payload || typeof payload !== 'object') return false;
+	const candidate = payload as Record<string, unknown>;
+	return (
+		typeof candidate.id === 'string' &&
+		typeof candidate.status === 'string' &&
+		WHISPER_JOB_STATUSES.has(candidate.status as WhisperJob['status'])
+	);
+}
+
 export async function POST({ params, request }) {
-const jobId = params.jobId;
-const job = getJob(jobId);
-if (!job) throw error(404, 'Job not found');
+	const jobId = params.jobId;
+	const job = getJob(jobId);
+	if (!job) throw error(404, 'Job not found');
 
-// Discard the result if the job was cancelled locally while whisper was running
-if (job.status === 'cancelled') {
-return json({ ok: true });
-}
-
-const whisperJob = (await request.json()) as WhisperJob;
-
-if (whisperJob.status === 'failed' || whisperJob.status === 'cancelled') {
-const msg = whisperJob.error ?? `Whisper job ${whisperJob.status}`;
-updateJob({ id: jobId, status: 'failed', error: msg });
-emitProgress(jobId, { type: 'error', message: msg });
-return json({ ok: true });
-}
-
-try {
-setJobStatus(jobId, 'processing', 90);
-emitProgress(jobId, { type: 'status', status: 'processing', progress: 90 });
-
-const rawSegments = whisperJob.segments as Segment[];
-const segments = deduplicateSegments(rawSegments);
-
-const paths = await writeOutputs(segments, job.title, jobId);
-const outputDir = paths.srt.replace(/\/[^/]+$/, '');
-
-updateJob({
-id: jobId,
-status: 'done',
-progress: 100,
-segmentsJson: JSON.stringify(segments),
-outputDir
-});
-
-emitProgress(jobId, { type: 'done', status: 'done' });
-
-await sendNotification(jobId, '✅ Transcript ready', job.title);
-await cleanupJobTmp(jobId);
-
-return json({ ok: true });
-} catch (err: unknown) {
-const message = err instanceof Error ? err.message : String(err);
-updateJob({ id: jobId, status: 'failed', error: message });
-emitProgress(jobId, { type: 'error', message });
-return json({ ok: false, error: message }, { status: 500 });
-}
+	const payload = (await request.json()) as unknown;
+	if (!isWhisperJobWebhook(payload)) {
+		// whisper-rtx2080 also fires model lifecycle events to registered job webhooks.
+		return json({ ok: true, ignored: 'not_a_job_event' });
+	}
+	const whisperJob = payload;
+
+	// Discard the result if the job was cancelled locally while whisper was running
+	if (job.status === 'cancelled') {
+		return json({ ok: true });
+	}
+
+	// Ignore stale callbacks from a previous whisper job after a local retry/reset.
+	if (job.whisperJobId && whisperJob.id !== job.whisperJobId) {
+		return json({ ok: true, ignored: 'stale_whisper_job' });
+	}
+
+	// Ignore replayed success callbacks after the transcript is already persisted.
+	if (job.status === 'done' && job.segmentsJson) {
+		return json({ ok: true, ignored: 'duplicate_webhook' });
+	}
+
+	if (whisperJob.status === 'failed' || whisperJob.status === 'cancelled') {
+		const msg = whisperJob.error ?? `Whisper job ${whisperJob.status}`;
+		updateJob({ id: jobId, status: 'failed', error: msg });
+		emitProgress(jobId, { type: 'error', message: msg });
+		return json({ ok: true });
+	}
+
+	try {
+		setJobStatus(jobId, 'processing', 90);
+		emitProgress(jobId, { type: 'status', status: 'processing', progress: 90 });
+
+		const segments = (whisperJob.segments ?? []) as Segment[];
+
+		const paths = await writeOutputs(segments, job.title, jobId);
+		const outputDir = paths.srt.replace(/\/[^/]+$/, '');
+
+		updateJob({
+			id: jobId,
+			status: 'done',
+			progress: 100,
+			segmentsJson: JSON.stringify(segments),
+			outputDir
+		});
+
+		emitProgress(jobId, { type: 'done', status: 'done' });
+
+		await sendNotification(jobId, '✅ Transcript ready', job.title);
+		await cleanupJobTmp(jobId);
+
+		return json({ ok: true });
+	} catch (err: unknown) {
+		const message = err instanceof Error ? err.message : String(err);
+		updateJob({ id: jobId, status: 'failed', error: message });
+		emitProgress(jobId, { type: 'error', message });
+		return json({ ok: false, error: message }, { status: 500 });
+	}
 }

src/routes/jobs/+page.svelte

@@ -1,6 +1,7 @@
 <script lang="ts">
 	import { onMount } from 'svelte';
 	import type { Job } from '$lib/types.js';
+	import { getDisplayJobProgress, getJobStatusColor, getJobStatusLabel, isTerminalJobStatus } from '$lib/job-progress.js';
 	import SourceIcon from '$lib/components/SourceIcon.svelte';
 	import Waveform from '$lib/components/Waveform.svelte';
 	import { accent } from '$lib/accent.js';
@@ -10,27 +11,6 @@
 	let jobs = $state<Job[]>([]);
 	let loading = $state(true);
 
-	const statusColor: Record<string, string> = {
-		done: '#cdf24e',
-		failed: '#ff6b6b',
-		cancelled: 'rgba(232,233,236,0.3)',
-		transcribing: '#80c7f7',
-		preparing: '#fbc94b',
-		downloading: '#a78bfa',
-		pending: 'rgba(232,233,236,0.4)'
-	};
-
-	const statusLabel: Record<string, string> = {
-		pending: 'Pending',
-		downloading: 'Downloading',
-		preparing: 'Preparing',
-		transcribing: 'Transcribing',
-		processing: 'Processing',
-		done: 'Done',
-		failed: 'Failed',
-		cancelled: 'Cancelled'
-	};
-
 	function jobKind(job: Job): 'youtube' | 'audio' | 'video' | 'file' {
 		const s = job.source ?? '';
 		if (s.includes('youtube') || s.includes('youtu.be')) return 'youtube';
@@ -108,8 +88,8 @@
 					<div class="job-info">
 						<div class="job-name">{job.title || job.source}</div>
 						<div class="job-meta mono">
-							<span style="color: {statusColor[job.status] ?? 'rgba(232,233,236,0.5)'}">
-								{statusLabel[job.status] ?? job.status}
+							<span style="color: {getJobStatusColor(job.status)}">
+								{getJobStatusLabel(job.status)}
 							</span>
 							{#if job.createdAt}
 								<span>·</span>
@@ -122,14 +102,24 @@
 						</div>
 					</div>
 
-					{#if !['done', 'failed', 'cancelled'].includes(job.status)}
+					{#if !isTerminalJobStatus(job.status)}
 						<div class="job-wave">
-							<Waveform bars={40} progress={job.progress} accent={ACCENT} height={28} pattern="medium" />
+							<Waveform
+								bars={40}
+								progress={getDisplayJobProgress(job, { hasTranscript: Boolean(job.segmentsJson) })}
+								accent={ACCENT}
+								height={28}
+								pattern="medium"
+							/>
 						</div>
 					{:else if job.status === 'done'}
-						<div class="job-pct mono" style="color: {ACCENT}">{job.progress}%</div>
+						<div class="job-pct mono" style="color: {ACCENT}">
+							{getDisplayJobProgress(job, { hasTranscript: Boolean(job.segmentsJson) })}%
+						</div>
 					{:else}
-						<div class="job-pct mono" style="color: {statusColor[job.status]}">{job.status}</div>
+						<div class="job-pct mono" style="color: {getJobStatusColor(job.status)}">
+							{getJobStatusLabel(job.status)}
+						</div>
 					{/if}
 
 					<!-- Row actions -->
@@ -142,7 +132,7 @@
 								title="Retry"
 							>↺</button>
 						{/if}
-						{#if ['done', 'failed', 'cancelled'].includes(job.status)}
+						{#if isTerminalJobStatus(job.status)}
 							<button
 								class="row-btn danger"
 								onclick={(e) => deleteJob(e, job)}

src/routes/jobs/[id]/+page.svelte

@@ -2,6 +2,7 @@
 	import { onMount, onDestroy } from 'svelte';
 	import { page } from '$app/stores';
 	import type { Job, Segment } from '$lib/types.js';
+	import { getDisplayJobProgress, getJobStatusLabel, isTerminalJobStatus } from '$lib/job-progress.js';
 	import SourceIcon from '$lib/components/SourceIcon.svelte';
 	import Waveform from '$lib/components/Waveform.svelte';
 	import { accent } from '$lib/accent.js';
@@ -16,17 +17,6 @@
 	let modelWarming = $state<{ state: string; retryAfterSecs: number } | null>(null);
 	let eventSource: EventSource | null = null;
 
-	const statusLabel: Record<string, string> = {
-		pending: 'Pending',
-		downloading: 'Downloading…',
-		preparing: 'Preparing audio…',
-		transcribing: 'Transcribing…',
-		processing: 'Post-processing…',
-		done: 'Done',
-		failed: 'Failed',
-		cancelled: 'Cancelled'
-	};
-
 	// Pipeline stages derived from job status
 	const pipelineStages = $derived.by(() => {
 		const status = job?.status ?? 'pending';
@@ -34,16 +24,26 @@
 			{ k: 'fetch', label: 'Fetch source' },
 			{ k: 'extract', label: 'Extract audio track' },
 			{ k: 'process', label: `Audio processing · ${job?.audioMode ?? 'auto'}` },
-			{ k: 'transcribe', label: 'Transcribing' },
+			{ k: 'transcribe', label: status === 'warming_model' ? 'Loading model' : 'Transcribing' },
 			{ k: 'finalize', label: 'Format &amp; save' }
 		];
-		const order = ['pending', 'downloading', 'preparing', 'transcribing', 'processing', 'done'];
-		const idx = order.indexOf(status);
+		const stageIndex = {
+			pending: 0,
+			downloading: 1,
+			preparing: 2,
+			warming_model: 3,
+			transcribing: 3,
+			processing: 4,
+			done: 5,
+			failed: -1,
+			cancelled: -1
+		}[status];
+
 		return stages.map((s, i) => ({
 			...s,
-			done: i < idx - 1 || status === 'done',
-			active: i === idx - 1 && status !== 'done' && status !== 'failed',
-			pending: i > idx - 1 && status !== 'done'
+			done: status === 'done' || i + 1 < stageIndex,
+			active: i + 1 === stageIndex && !isTerminalJobStatus(status),
+			pending: status !== 'done' && i + 1 > stageIndex
 		}));
 	});
 
@@ -57,7 +57,7 @@
 
 	onMount(async () => {
 		await loadJob();
-		if (job && !['done', 'failed', 'cancelled'].includes(job.status)) {
+		if (job && !isTerminalJobStatus(job.status)) {
 			openStream();
 		}
 	});
@@ -71,6 +71,7 @@
 			return;
 		}
 		job = await res.json();
+		segments = [];
 		if (job?.segmentsJson) {
 			try {
 				segments = JSON.parse(job.segmentsJson);
@@ -89,9 +90,14 @@
 					if (job) job = { ...job, progress: data.progress ?? job.progress, status: 'transcribing' };
 				} else if (data.type === 'model_warming') {
 					modelWarming = { state: data.state ?? 'loading', retryAfterSecs: data.retryAfterSecs ?? 30 };
+					chunkInfo = { chunk: 0, total: 0 };
+					if (job) job = { ...job, status: 'warming_model', progress: data.progress ?? job.progress };
 				} else if (data.type === 'status') {
+					if (data.status !== 'warming_model') modelWarming = null;
+					if (data.status !== 'transcribing') chunkInfo = { chunk: 0, total: 0 };
 					if (job) job = { ...job, status: data.status, progress: data.progress ?? job.progress };
 				} else if (data.type === 'done') {
+					modelWarming = null;
 					eventSource?.close();
 					loadJob();
 				} else if (data.type === 'error') {
@@ -112,8 +118,21 @@
 	}
 
 	const formats = ['srt', 'txt', 'md', 'json'] as const;
-	const isActive = $derived(!job || !['done', 'failed', 'cancelled'].includes(job.status));
-	const isTerminal = $derived(job !== null && ['done', 'failed', 'cancelled'].includes(job.status));
+	const hasTranscript = $derived((job?.segmentsJson ? true : false) || segments.length > 0);
+	const displayProgress = $derived(job ? getDisplayJobProgress(job, { hasTranscript }) : 0);
+	const progressStatusLabel = $derived.by(() => {
+		if (!job) return 'Pending';
+		if (job.status === 'warming_model') {
+			const state = modelWarming?.state?.replace(/_/g, ' ');
+			return state ? `Loading model (${state})…` : 'Loading model…';
+		}
+		if (job.status === 'preparing') return 'Preparing audio…';
+		if (job.status === 'processing') return 'Saving transcript…';
+		if (job.status === 'transcribing') return 'Transcribing…';
+		return job.status === 'done' ? 'Done' : `${getJobStatusLabel(job.status)}…`;
+	});
+	const isActive = $derived(!job || !isTerminalJobStatus(job.status));
+	const isTerminal = $derived(job !== null && isTerminalJobStatus(job.status));
 	const canRetry = $derived(
 		job !== null &&
 		['failed', 'cancelled'].includes(job.status) &&
@@ -198,7 +217,7 @@
 				<div class="progress-wave">
 						<Waveform
 							bars={140}
-						progress={job.progress}
+							progress={displayProgress}
 							accent={ACCENT}
 							height={80}
 							pattern="default"
@@ -208,9 +227,9 @@
 					<div class="progress-footer">
 						<div class="progress-left">
 							<span class="progress-pct mono">
-							{job.progress}<span style="color: var(--text-dim); font-weight: 400">%</span>
+							{displayProgress}<span style="color: var(--text-dim); font-weight: 400">%</span>
 							</span>
-						<span class="progress-status">{statusLabel[job.status] ?? job.status}</span>
+						<span class="progress-status">{progressStatusLabel}</span>
 						</div>
 					{#if chunkInfo.total > 0}
 						<span class="progress-chunks mono">
@@ -232,7 +251,7 @@
 				<div class="progress-bar-track">
 					<div
 						class="progress-bar-fill"
-						style="width: {job.progress}%; background: {ACCENT}; box-shadow: 0 0 12px {ACCENT}80;"
+						style="width: {displayProgress}%; background: {ACCENT}; box-shadow: 0 0 12px {ACCENT}80;"
 					></div>
 				</div>
 			</div>
@@ -274,7 +293,7 @@
 								{@html stage.label}
 							</span>
 							{#if stage.active}
-								<span class="mono" style="font-size: 11.5px; color: {ACCENT}">{job.progress}%</span>
+								<span class="mono" style="font-size: 11.5px; color: {ACCENT}">{displayProgress}%</span>
 							{/if}
 						</div>
 					{/each}
@@ -689,4 +708,3 @@
 		}
 	}
 </style>
-

src/tests/audio.test.ts

@@ -7,17 +7,18 @@ import { describe, it, expect, vi, afterEach } from 'vitest';
 
 const execFileMock = vi.hoisted(() => {
 	const fn = vi.fn();
+	type ExecFilePromisifyArgs = [string, string[]];
+	type ExecFileCallback = (err: Error | null, stdout: string, stderr: string) => void;
+	type ExecFileMock = (...args: [...ExecFilePromisifyArgs, ExecFileCallback]) => void;
+	const invoke = fn as unknown as ExecFileMock;
 	Object.defineProperty(fn, Symbol.for('nodejs.util.promisify.custom'), {
 		configurable: true,
-		value: (...args: unknown[]) =>
+		value: (...args: ExecFilePromisifyArgs) =>
 			new Promise<{ stdout: string; stderr: string }>((resolve, reject) => {
-				(fn as ReturnType<typeof vi.fn>)(
-					...args,
-					(err: Error | null, stdout: string, stderr: string) => {
+				invoke(...args, (err: Error | null, stdout: string, stderr: string) => {
 					if (err) reject(err);
 					else resolve({ stdout, stderr });
-					}
-				);
+				});
 			})
 	});
 	return fn;

src/tests/db.test.ts

@@ -123,7 +123,7 @@ describe('setJobStatus', () => {
 
 	it('transitions through all valid statuses', () => {
 		const job = createJob('src', 'title', 'auto');
-		const statuses = ['downloading', 'preparing', 'transcribing', 'processing', 'done'] as const;
+		const statuses = ['downloading', 'preparing', 'warming_model', 'transcribing', 'processing', 'done'] as const;
 		for (const status of statuses) {
 			setJobStatus(job.id, status, 50);
 			expect(getJob(job.id)!.status).toBe(status);

src/tests/downloader.test.ts

@@ -0,0 +1,80 @@
+import { afterEach, beforeEach, describe, expect, it, vi } from 'vitest';
+import { rm } from 'fs/promises';
+import type { TranscriptResponse } from 'youtube-transcript';
+
+const { mockExecFile, mockFetchTranscript } = vi.hoisted(() => ({
+	mockExecFile: vi.fn(),
+	mockFetchTranscript: vi.fn()
+}));
+
+const TEST_DATA_DIR = `/tmp/tonemark-downloader-test-${Date.now()}`;
+vi.stubEnv('DATA_DIR', TEST_DATA_DIR);
+
+vi.mock('child_process', () => ({
+	execFile: mockExecFile
+}));
+
+vi.mock('youtube-transcript', () => ({
+	fetchTranscript: mockFetchTranscript
+}));
+
+import { downloadYouTube, transcriptEntriesToSegments } from '$lib/server/downloader.js';
+
+beforeEach(() => {
+	vi.clearAllMocks();
+	mockExecFile.mockImplementation((...args: unknown[]) => {
+		const cb = args.at(-1) as (...callbackArgs: unknown[]) => void;
+		cb(null, JSON.stringify({ title: 'Fetched Title' }), '');
+	});
+});
+
+afterEach(async () => {
+	await rm(TEST_DATA_DIR, { recursive: true, force: true }).catch(() => {});
+});
+
+describe('transcriptEntriesToSegments', () => {
+	it('converts millisecond transcript offsets into second-based segments', () => {
+		const entries: TranscriptResponse[] = [
+			{ text: 'Hello everyone.', offset: 15240, duration: 4240, lang: 'en' },
+			{ text: 'Um, welcome to this talk.', offset: 16600, duration: 5080, lang: 'en' }
+		];
+
+		expect(transcriptEntriesToSegments(entries)).toEqual([
+			{ index: 0, start: 15.24, end: 19.48, text: 'Hello everyone.', words: [] },
+			{ index: 1, start: 16.6, end: 21.68, text: 'Um, welcome to this talk.', words: [] }
+		]);
+	});
+
+	it('preserves second-based transcript offsets and drops empty text', () => {
+		const entries: TranscriptResponse[] = [
+			{ text: '  ', offset: 0, duration: 1.5, lang: 'en' },
+			{ text: 'Clean caption cue', offset: 91.08, duration: 3.72, lang: 'en' }
+		];
+
+		expect(transcriptEntriesToSegments(entries)).toEqual([
+			{ index: 0, start: 91.08, end: 94.8, text: 'Clean caption cue', words: [] }
+		]);
+	});
+});
+
+describe('downloadYouTube', () => {
+	it('uses fetched transcript entries directly for caption jobs', async () => {
+		mockFetchTranscript.mockResolvedValue([
+			{ text: 'Hello everyone.', offset: 15240, duration: 4240, lang: 'en' },
+			{ text: 'Um, welcome to this talk.', offset: 16600, duration: 5080, lang: 'en' }
+		] satisfies TranscriptResponse[]);
+
+		const result = await downloadYouTube('https://youtube.com/watch?v=qdh_x-uRs9g', 'job-1');
+
+		expect(mockFetchTranscript).toHaveBeenCalledWith('https://youtube.com/watch?v=qdh_x-uRs9g', {
+			lang: 'en'
+		});
+		expect(result).toMatchObject({
+			type: 'captions',
+			segments: [
+				{ index: 0, start: 15.24, end: 19.48, text: 'Hello everyone.', words: [] },
+				{ index: 1, start: 16.6, end: 21.68, text: 'Um, welcome to this talk.', words: [] }
+			]
+		});
+	});
+});

src/tests/job-progress.test.ts

@@ -0,0 +1,49 @@
+import { describe, expect, it } from 'vitest';
+import { getDisplayJobProgress, getJobStatusLabel, isTerminalJobStatus } from '$lib/job-progress.js';
+import type { Job } from '$lib/types.js';
+
+function makeJob(overrides: Partial<Job> = {}): Job {
+	return {
+		id: 'job-1',
+		status: 'transcribing',
+		title: 'Job',
+		source: 'https://example.com/audio.mp3',
+		audioMode: 'auto',
+		meanVolume: null,
+		whisperJobId: null,
+		progress: 42,
+		outputDir: null,
+		segmentsJson: null,
+		error: null,
+		createdAt: new Date().toISOString(),
+		updatedAt: new Date().toISOString(),
+		...overrides
+	};
+}
+
+describe('job progress helpers', () => {
+	it('keeps active jobs below 100 percent', () => {
+		expect(getDisplayJobProgress(makeJob({ status: 'transcribing', progress: 100 }))).toBe(99);
+	});
+
+	it('keeps model loading in an early progress band', () => {
+		expect(getDisplayJobProgress(makeJob({ status: 'warming_model', progress: 80 }))).toBe(15);
+	});
+
+	it('allows 100 percent once finished job has transcript payload', () => {
+		expect(
+			getDisplayJobProgress(makeJob({ status: 'done', progress: 100, segmentsJson: JSON.stringify([]) }), {
+				hasTranscript: true
+			})
+		).toBe(100);
+	});
+
+	it('holds done jobs below 100 percent until transcript data exists', () => {
+		expect(getDisplayJobProgress(makeJob({ status: 'done', progress: 100 }))).toBe(99);
+	});
+
+	it('exposes model-loading label as active state', () => {
+		expect(getJobStatusLabel('warming_model')).toBe('Loading model');
+		expect(isTerminalJobStatus('warming_model')).toBe(false);
+	});
+});

src/tests/postprocess.test.ts

@@ -1,127 +0,0 @@
-import { describe, it, expect } from 'vitest';
-import {
-	deduplicateSegments
-} from '$lib/server/postprocess.js';
-import type { Segment } from '$lib/types.js';
-
-// ── helpers ──────────────────────────────────────────────────────────────────
-
-function seg(index: number, start: number, end: number, text: string): Segment {
-	return { index, start, end, text, words: [] };
-}
-
-// ── collapseRepeats (tested indirectly via deduplicateSegments) ───────────────
-
-describe('deduplicateSegments — collapseRepeats', () => {
-	it('leaves text without repetition unchanged', () => {
-		const input = [seg(0, 0, 5, ' Hello world, this is a sentence.')];
-		const [out] = deduplicateSegments(input);
-		expect(out.text).toBe('Hello world, this is a sentence.');
-	});
-
-	it('collapses a consecutive repeated phrase inside a segment', () => {
-		const input = [seg(0, 0, 5, ' the quick brown fox the quick brown fox')];
-		const [out] = deduplicateSegments(input);
-		expect(out.text).not.toMatch(/the quick brown fox.*the quick brown fox/i);
-	});
-
-	it('handles multiple repetitions recursively', () => {
-		// "welcome everyone" = 16 chars — qualifies for the ≥10-char collapse regex
-		const input = [seg(0, 0, 5, ' welcome everyone welcome everyone welcome everyone')];
-		const result = deduplicateSegments(input);
-		const text = result[0]?.text ?? '';
-		expect((text.match(/welcome everyone/gi) ?? []).length).toBeLessThan(3);
-	});
-});
-
-// ── mergeConsecutive ──────────────────────────────────────────────────────────
-
-describe('deduplicateSegments — mergeConsecutive', () => {
-	it('merges adjacent segments with identical text', () => {
-		const input = [
-			seg(0, 0, 2, ' Hello world.'),
-			seg(1, 2, 4, ' Hello world.')
-		];
-		const result = deduplicateSegments(input);
-		expect(result).toHaveLength(1);
-		expect(result[0].end).toBe(4);
-	});
-
-	it('keeps adjacent segments with different text', () => {
-		const input = [
-			seg(0, 0, 2, ' First sentence.'),
-			seg(1, 2, 4, ' Second sentence.')
-		];
-		const result = deduplicateSegments(input);
-		expect(result).toHaveLength(2);
-	});
-
-	it('normalises punctuation and case for merge comparison', () => {
-		const input = [
-			seg(0, 0, 2, ' Hello, World!'),
-			seg(1, 2, 4, ' hello world')
-		];
-		const result = deduplicateSegments(input);
-		expect(result).toHaveLength(1);
-	});
-});
-
-// ── ngramDedup ────────────────────────────────────────────────────────────────
-
-describe('deduplicateSegments — ngramDedup', () => {
-	it('passes through completely unique segments', () => {
-		const input = [
-			seg(0, 0, 5, ' The cat sat on the mat quite happily today.'),
-			seg(1, 5, 10, ' Later the dog ran across the yard chasing a ball.')
-		];
-		expect(deduplicateSegments(input)).toHaveLength(2);
-	});
-
-	it('removes a segment that is highly similar to recent context', () => {
-		// Repeat a long sentence verbatim — should be caught as duplicate
-		const longText =
-			' This is a very specific and unique sentence about transcription quality matters greatly.';
-		const input = [seg(0, 0, 5, longText), seg(1, 5, 10, longText)];
-		// After mergeConsecutive the second one is already merged, so result is 1
-		expect(deduplicateSegments(input)).toHaveLength(1);
-	});
-});
-
-// ── deduplicateSegments — full pipeline ──────────────────────────────────────
-
-describe('deduplicateSegments — full pipeline', () => {
-	it('returns empty array for empty input', () => {
-		expect(deduplicateSegments([])).toEqual([]);
-	});
-
-	it('removes segments whose text is empty after trimming', () => {
-		const input = [seg(0, 0, 1, '   '), seg(1, 1, 2, ' Hello.')];
-		const result = deduplicateSegments(input);
-		expect(result).toHaveLength(1);
-		expect(result[0].text).toBe('Hello.');
-	});
-
-	it('re-indexes output segments starting from 0', () => {
-		const input = [
-			seg(5, 0, 2, ' First unique sentence here.'),
-			seg(8, 2, 4, ' Second different sentence there.')
-		];
-		const result = deduplicateSegments(input);
-		result.forEach((s, i) => expect(s.index).toBe(i));
-	});
-
-	it('runs the full pipeline: trim → remove empty → merge → ngram → merge → reindex', () => {
-		const input = [
-			seg(0, 0, 2, ' Good morning everyone.'),
-			seg(1, 2, 3, '   '), // empty — removed
-			seg(2, 3, 5, ' Good morning everyone.'), // duplicate — merged
-			seg(3, 5, 7, ' Welcome to our presentation today.')
-		];
-		const result = deduplicateSegments(input);
-		expect(result).toHaveLength(2);
-		expect(result[0].text).toBe('Good morning everyone.');
-		expect(result[1].text).toBe('Welcome to our presentation today.');
-		expect(result[0].index).toBe(0);
-		expect(result[1].index).toBe(1);
-	});
-});

src/tests/push.test.ts

@@ -1,4 +1,4 @@
-import { describe, it, expect, vi, afterEach } from 'vitest';
+import { describe, it, expect, vi, beforeEach, afterEach } from 'vitest';
 
 // ── Hoist mock functions so they're available inside vi.mock() factories ───────
 const { mockSetVapidDetails, mockWebPushSend } = vi.hoisted(() => ({
@@ -24,12 +24,16 @@ import { sendNotification, getVapidPublicKey } from '$lib/server/push.js';
 import { savePushSubscription, deletePushSubscription, getAllSubscriptions } from '$lib/server/db.js';
 import { rm } from 'fs/promises';
 
-afterEach(async () => {
-	mockSetVapidDetails.mockReset();
+beforeEach(() => {
+	// Ensure a clean subscription table before each test
+	for (const s of getAllSubscriptions()) deletePushSubscription(s.endpoint);
 	mockWebPushSend.mockReset();
+	mockSetVapidDetails.mockReset();
+});
+
+afterEach(async () => {
 	// Remove all test subscriptions between tests
-	const subs = getAllSubscriptions();
-	for (const s of subs) deletePushSubscription(s.endpoint);
+	for (const s of getAllSubscriptions()) deletePushSubscription(s.endpoint);
 	await rm(TEST_DATA_DIR, { recursive: true, force: true }).catch(() => {});
 });
 
@@ -134,6 +138,15 @@ describe('sendNotification', () => {
 			.mockResolvedValueOnce({});
 
 		await sendNotification('job-8', 'title', 'body');
-		expect(mockWebPushSend).toHaveBeenCalledTimes(3);
+		const calledEndpoints = mockWebPushSend.mock.calls.map(
+			([sub]) => (sub as { endpoint: string }).endpoint
+		);
+		expect(calledEndpoints).toEqual(
+			expect.arrayContaining([
+				'https://fcm.example.com/push/ok1',
+				'https://fcm.example.com/push/fail',
+				'https://fcm.example.com/push/ok2'
+			])
+		);
 	});
 });

src/tests/webhook.test.ts

@@ -7,7 +7,6 @@ const {
 	mockGetJob,
 	mockUpdateJob,
 	mockSetJobStatus,
-	mockDeduplicateSegments,
 	mockWriteOutputs,
 	mockSendNotification,
 	mockCleanupJobTmp,
@@ -16,7 +15,6 @@ const {
 	mockGetJob: vi.fn(),
 	mockUpdateJob: vi.fn(),
 	mockSetJobStatus: vi.fn(),
-	mockDeduplicateSegments: vi.fn((segs: Segment[]) => segs),
 	mockWriteOutputs: vi.fn(),
 	mockSendNotification: vi.fn(),
 	mockCleanupJobTmp: vi.fn(),
@@ -29,10 +27,6 @@ vi.mock('$lib/server/db.js', () => ({
 	setJobStatus: mockSetJobStatus
 }));
 
-vi.mock('$lib/server/postprocess.js', () => ({
-	deduplicateSegments: mockDeduplicateSegments
-}));
-
 vi.mock('$lib/server/formatter.js', () => ({
 	writeOutputs: mockWriteOutputs
 }));
@@ -91,7 +85,6 @@ function makeSeg(index: number, text: string): Segment {
 
 beforeEach(() => {
 	vi.clearAllMocks();
-	mockDeduplicateSegments.mockImplementation((segs: Segment[]) => segs);
 	mockWriteOutputs.mockResolvedValue({
 		srt: '/out/dir/title.srt',
 		txt: '/out/dir/title.txt',
@@ -113,6 +106,64 @@ describe('POST /api/webhook/[jobId] — job not found', () => {
 	});
 });
 
+// ── Ignore backend model lifecycle webhooks ─────────────────────────────────────
+
+describe('POST /api/webhook/[jobId] — non-job webhook payloads', () => {
+	it('ignores model_ready events sent to job webhooks', async () => {
+		mockGetJob.mockReturnValue(makeJob('job-model-ready'));
+
+		const res = await POST(
+			makeEvent('job-model-ready', {
+				type: 'model_ready',
+				loaded_at: new Date().toISOString()
+			}) as any
+		);
+
+		expect(res.status).toBe(200);
+		expect(await res.json()).toEqual({ ok: true, ignored: 'not_a_job_event' });
+		expect(mockSetJobStatus).not.toHaveBeenCalled();
+		expect(mockUpdateJob).not.toHaveBeenCalled();
+		expect(mockWriteOutputs).not.toHaveBeenCalled();
+		expect(mockSendNotification).not.toHaveBeenCalled();
+	});
+
+	it('ignores model_unloaded events sent to job webhooks', async () => {
+		mockGetJob.mockReturnValue(makeJob('job-model-unloaded'));
+
+		const res = await POST(
+			makeEvent('job-model-unloaded', {
+				type: 'model_unloaded',
+				unloaded_at: new Date().toISOString()
+			}) as any
+		);
+
+		expect(res.status).toBe(200);
+		expect(await res.json()).toEqual({ ok: true, ignored: 'not_a_job_event' });
+		expect(mockSetJobStatus).not.toHaveBeenCalled();
+		expect(mockUpdateJob).not.toHaveBeenCalled();
+		expect(mockWriteOutputs).not.toHaveBeenCalled();
+		expect(mockSendNotification).not.toHaveBeenCalled();
+	});
+
+	it('ignores payloads with invalid status values', async () => {
+		mockGetJob.mockReturnValue(makeJob('job-invalid-status'));
+
+		const res = await POST(
+			makeEvent('job-invalid-status', {
+				id: 'bogus-whisper-id',
+				status: 'model_ready',
+				segments: []
+			}) as any
+		);
+
+		expect(res.status).toBe(200);
+		expect(await res.json()).toEqual({ ok: true, ignored: 'not_a_job_event' });
+		expect(mockSetJobStatus).not.toHaveBeenCalled();
+		expect(mockUpdateJob).not.toHaveBeenCalled();
+		expect(mockWriteOutputs).not.toHaveBeenCalled();
+	});
+});
+
 // ── Local cancellation guard ──────────────────────────────────────────────────
 
 describe('POST /api/webhook/[jobId] — locally cancelled job', () => {
@@ -132,6 +183,43 @@ describe('POST /api/webhook/[jobId] — locally cancelled job', () => {
 	});
 });
 
+// ── Duplicate / stale callback guards ──────────────────────────────────────────
+
+describe('POST /api/webhook/[jobId] — duplicate and stale callbacks', () => {
+	it('ignores replayed success callbacks after the transcript is already done', async () => {
+		mockGetJob.mockReturnValue({
+			...makeJob('job-done'),
+			status: 'done',
+			segmentsJson: JSON.stringify([makeSeg(0, 'Already saved.')]),
+			whisperJobId: 'whisper-id'
+		});
+
+		const res = await POST(makeEvent('job-done', makeWhisperJob()) as any);
+		expect(res.status).toBe(200);
+		expect(await res.json()).toEqual({ ok: true, ignored: 'duplicate_webhook' });
+		expect(mockSetJobStatus).not.toHaveBeenCalled();
+		expect(mockUpdateJob).not.toHaveBeenCalled();
+		expect(mockWriteOutputs).not.toHaveBeenCalled();
+	});
+
+	it('ignores stale callbacks from an older whisper job after retry', async () => {
+		mockGetJob.mockReturnValue({
+			...makeJob('job-stale'),
+			status: 'transcribing',
+			whisperJobId: 'current-whisper-job'
+		});
+
+		const res = await POST(
+			makeEvent('job-stale', makeWhisperJob({ id: 'old-whisper-job', segments: [makeSeg(0, 'stale')] })) as any
+		);
+		expect(res.status).toBe(200);
+		expect(await res.json()).toEqual({ ok: true, ignored: 'stale_whisper_job' });
+		expect(mockSetJobStatus).not.toHaveBeenCalled();
+		expect(mockUpdateJob).not.toHaveBeenCalled();
+		expect(mockWriteOutputs).not.toHaveBeenCalled();
+	});
+});
+
 // ── Whisper job failed / cancelled ───────────────────────────────────────────
 
 describe('POST /api/webhook/[jobId] — whisper failure', () => {
@@ -181,25 +269,21 @@ describe('POST /api/webhook/[jobId] — whisper failure', () => {
 describe('POST /api/webhook/[jobId] — success with segments', () => {
 	const segments = [makeSeg(0, 'Hello world.'), makeSeg(1, 'This is a test.')];
 
-	it('runs deduplication on received segments', async () => {
+	it('passes received segments through unchanged', async () => {
 		mockGetJob.mockReturnValue(makeJob('job-3'));
 		await POST(makeEvent('job-3', makeWhisperJob({ segments })) as any);
-		expect(mockDeduplicateSegments).toHaveBeenCalledWith(segments);
+		expect(mockWriteOutputs).toHaveBeenCalledWith(segments, 'Test Video', 'job-3');
 	});
 
-	it('calls writeOutputs with the deduplicated segments and job title', async () => {
+	it('calls writeOutputs with the received segments and job title', async () => {
 		mockGetJob.mockReturnValue(makeJob('job-4', 'My Lecture'));
-		const deduped = [makeSeg(0, 'Hello world.')];
-		mockDeduplicateSegments.mockReturnValue(deduped);
 
 		await POST(makeEvent('job-4', makeWhisperJob({ segments })) as any);
-		expect(mockWriteOutputs).toHaveBeenCalledWith(deduped, 'My Lecture', 'job-4');
+		expect(mockWriteOutputs).toHaveBeenCalledWith(segments, 'My Lecture', 'job-4');
 	});
 
 	it('stores serialised segments_json in the database', async () => {
 		mockGetJob.mockReturnValue(makeJob('job-5'));
-		const deduped = [makeSeg(0, 'Result text.')];
-		mockDeduplicateSegments.mockReturnValue(deduped);
 
 		await POST(makeEvent('job-5', makeWhisperJob({ segments })) as any);
 
@@ -207,7 +291,7 @@ describe('POST /api/webhook/[jobId] — success with segments', () => {
 			expect.objectContaining({
 				id: 'job-5',
 				status: 'done',
-				segmentsJson: JSON.stringify(deduped)
+				segmentsJson: JSON.stringify(segments)
 			})
 		);
 	});
@@ -289,6 +373,34 @@ describe('POST /api/webhook/[jobId] — empty segments', () => {
 	});
 });
 
+// ── Undefined / missing segments (model returned no segments field) ───────────
+
+describe('POST /api/webhook/[jobId] — undefined segments', () => {
+	it('completes the job as done when segments field is absent from whisper payload', async () => {
+		mockGetJob.mockReturnValue(makeJob('job-noseg'));
+		// Simulate whisper returning a result without a segments field
+		const payload = { ...makeWhisperJob(), segments: undefined as unknown as never[] };
+
+		const res = await POST(makeEvent('job-noseg', payload) as any);
+		expect(res.status).toBe(200);
+		expect(await res.json()).toEqual({ ok: true });
+		expect(mockUpdateJob).toHaveBeenCalledWith(
+			expect.objectContaining({ status: 'done', id: 'job-noseg' })
+		);
+	});
+
+	it('does not throw "cannot read properties of undefined" when segments is null', async () => {
+		mockGetJob.mockReturnValue(makeJob('job-nullseg'));
+		const payload = { ...makeWhisperJob(), segments: null as unknown as never[] };
+
+		// Must NOT throw — previously crashed with "Cannot read properties of undefined (reading 'map')"
+		await expect(POST(makeEvent('job-nullseg', payload) as any)).resolves.toBeDefined();
+		expect(mockUpdateJob).toHaveBeenCalledWith(
+			expect.objectContaining({ status: 'done' })
+		);
+	});
+});
+
 // ── Internal error handling ───────────────────────────────────────────────────
 
 describe('POST /api/webhook/[jobId] — internal errors', () => {

src/tests/whisper.test.ts

@@ -6,7 +6,8 @@ import { Readable } from 'stream';
 const mocks = vi.hoisted(() => ({
 	fetch: vi.fn(),
 	append: vi.fn(),
-	getHeaders: vi.fn(() => ({ 'content-type': 'multipart/form-data; boundary=test' }))
+	getHeaders: vi.fn(() => ({ 'content-type': 'multipart/form-data; boundary=test' })),
+	createReadStream: vi.fn(() => 'STREAM_PLACEHOLDER')
 }));
 
 vi.mock('node-fetch', () => ({ default: mocks.fetch }));
@@ -19,9 +20,9 @@ vi.mock('form-data', () => ({
 	})
 }));
 
-vi.mock('fs', () => ({ createReadStream: vi.fn(() => 'STREAM_PLACEHOLDER') }));
+vi.mock('fs', () => ({ createReadStream: mocks.createReadStream }));
 
-import { submitJob, streamJob, getModelStatus, cancelJob } from '$lib/server/whisper.js';
+import { submitJob, streamJob, getModelStatus, cancelJob, unloadModel } from '$lib/server/whisper.js';
 
 afterEach(() => vi.clearAllMocks());
 
@@ -255,6 +256,20 @@ describe('submitJob — 503 retry behavior', () => {
 		expect(onModelWaiting).not.toHaveBeenCalled();
 	});
 
+	it('creates a fresh ReadStream for each attempt (stream not reused across retries)', async () => {
+		mocks.fetch.mockImplementation(
+			makeJobFetch(make503('loading', 0), make503('loading', 0), make202('fresh-stream-id'))
+		);
+
+		const id = await submitJob('/tmp/audio.wav', 'http://host/webhook', undefined, undefined, 10);
+		expect(id).toBe('fresh-stream-id');
+		// 3 attempts → 3 separate createReadStream calls, one fresh stream per form
+		expect(mocks.createReadStream).toHaveBeenCalledTimes(3);
+		expect(mocks.createReadStream).toHaveBeenNthCalledWith(1, '/tmp/audio.wav');
+		expect(mocks.createReadStream).toHaveBeenNthCalledWith(2, '/tmp/audio.wav');
+		expect(mocks.createReadStream).toHaveBeenNthCalledWith(3, '/tmp/audio.wav');
+	});
+
 	it('does NOT retry on non-503 errors (throws immediately)', async () => {
 		mocks.fetch.mockResolvedValue({
 			status: 400,
@@ -339,6 +354,68 @@ describe('submitJob — SSE-triggered retry', () => {
 		const id = await submitJob('/tmp/audio.wav', 'http://host/webhook');
 		expect(id).toBe('stream-closed-id');
 	});
+
+	it('relays intermediate model states from /model/events while waiting to retry', async () => {
+		let jobCallIdx = 0;
+		mocks.fetch.mockImplementation((url: string) => {
+			if (String(url).includes('/model/events')) {
+				return Promise.resolve({
+					ok: true,
+					status: 200,
+					body: Readable.from([
+						'data: {"state":"loading"}\n\ndata: {"state":"waiting_for_gpu"}\n\ndata: {"state":"ready"}\n\n'
+					])
+				});
+			}
+			jobCallIdx++;
+			if (jobCallIdx === 1) return Promise.resolve(make503('unloaded', 30));
+			return Promise.resolve(make202('state-relay-id'));
+		});
+
+		const onModelWaiting = vi.fn();
+		const id = await submitJob('/tmp/audio.wav', 'http://host/webhook', undefined, onModelWaiting);
+		expect(id).toBe('state-relay-id');
+		expect(onModelWaiting).toHaveBeenNthCalledWith(1, 'unloaded', 30);
+		expect(onModelWaiting).toHaveBeenNthCalledWith(2, 'loading', 30);
+		expect(onModelWaiting).toHaveBeenNthCalledWith(3, 'waiting_for_gpu', 30);
+		expect(onModelWaiting).toHaveBeenCalledTimes(3);
+	});
+});
+
+// ── unloadModel ───────────────────────────────────────────────────────────────
+
+describe('unloadModel', () => {
+	it('POSTs to /model/unload and returns parsed body', async () => {
+		mocks.fetch.mockResolvedValue({
+			ok: true,
+			json: () => Promise.resolve({ ok: true })
+		});
+		const result = await unloadModel();
+		expect(result).toEqual({ ok: true });
+		expect(mocks.fetch).toHaveBeenCalledWith(
+			expect.stringContaining('/model/unload'),
+			expect.objectContaining({ method: 'POST' })
+		);
+	});
+
+	it('uses the configured WHISPER_URL', async () => {
+		vi.stubEnv('WHISPER_URL', 'http://gpu-box:9090');
+		mocks.fetch.mockResolvedValue({
+			ok: true,
+			json: () => Promise.resolve({ ok: true })
+		});
+		await unloadModel();
+		expect(mocks.fetch).toHaveBeenCalledWith(
+			'http://gpu-box:9090/model/unload',
+			expect.anything()
+		);
+		vi.unstubAllEnvs();
+	});
+
+	it('throws when whisper returns a non-ok response', async () => {
+		mocks.fetch.mockResolvedValue({ ok: false, status: 409 });
+		await expect(unloadModel()).rejects.toThrow('/model/unload');
+	});
 });
 
 // ── cancelJob ─────────────────────────────────────────────────────────────────
@@ -458,6 +535,10 @@ function makeSSEResponse(lines: string[]) {
 	return { ok: true, body };
 }
 
+function makeSSEChunkResponse(chunks: string[]) {
+	return { ok: true, body: Readable.from(chunks) };
+}
+
 describe('streamJob — SSE event parsing', () => {
 	it('calls onProgress for progress events with percent, chunk, total', async () => {
 		const onProgress = vi.fn();
@@ -551,6 +632,27 @@ describe('streamJob — SSE event parsing', () => {
 		expect(onProgress).toHaveBeenNthCalledWith(3, 75, 3, 4);
 	});
 
+	it('handles multiple SSE events delivered in a single chunk', async () => {
+		const onProgress = vi.fn();
+		const onDone = vi.fn();
+		const onError = vi.fn();
+
+		mocks.fetch.mockResolvedValue(
+			makeSSEChunkResponse([
+				'data: {"type":"progress","percent":25,"chunk":1,"total":2}\n\n' +
+				'data: {"type":"progress","percent":50,"chunk":2,"total":2}\n\n' +
+				'data: {"type":"done","job":{}}\n\n'
+			])
+		);
+
+		await streamJob('whisper-id', onProgress, onDone, onError);
+		expect(onProgress).toHaveBeenCalledTimes(2);
+		expect(onProgress).toHaveBeenNthCalledWith(1, 25, 1, 2);
+		expect(onProgress).toHaveBeenNthCalledWith(2, 50, 2, 2);
+		expect(onDone).toHaveBeenCalledOnce();
+		expect(onError).not.toHaveBeenCalled();
+	});
+
 	it('defaults chunk and total to 0 when missing from progress event', async () => {
 		const onProgress = vi.fn();
 		const onDone = vi.fn();

vitest.config.ts

@@ -5,6 +5,7 @@ export default defineConfig({
 	test: {
 		environment: 'node',
 		globals: true,
+		fileParallelism: false,
 		include: ['src/tests/**/*.test.ts'],
 		coverage: {
 			provider: 'v8',