fix(stt): distinguish initializing from transcribing on every spinner

src/components/ai-edition/CaptionsPane.gating.test.tsx

@@ -107,7 +107,36 @@ describe("captions pane gating", () => {
 				<CaptionsPane />
 			</I18nProvider>,
 		);
-		expect(screen.getByRole("button", { name: "Transcribing…" })).toBeDisabled();
+		// A phase-less running job renders the shared busy label ("Transcribing",
+		// mediaStage.transcribing) rather than the pane's old private copy.
+		expect(screen.getByRole("button", { name: "Transcribing" })).toBeDisabled();
+	});
+
+	it("keeps the idle button when only an off-timeline asset is busy", () => {
+		// The gate answers for the timeline's assets; the label must not answer
+		// for the whole bin. A bin asset mid-transcription used to relabel the
+		// still-enabled button with its busy copy.
+		const offTimeline: AxcutAsset = {
+			id: "asset_2",
+			kind: "video",
+			label: "bin-only.mp4",
+			originalPath: "/bin.mp4",
+			durationSec: 8,
+			cameraTrack: null,
+		};
+		const document = documentWith(ASSET);
+		document.assets.push(offTimeline);
+		load(document);
+		useTranscriptionStore.setState({
+			projectId: "proj_1",
+			jobs: { asset_2: { status: "running", language: "auto", manual: false } },
+		});
+		render(
+			<I18nProvider>
+				<CaptionsPane />
+			</I18nProvider>,
+		);
+		expect(screen.getByRole("button", { name: "Transcribe video" })).toBeEnabled();
 	});
 
 	it("kills the retry on a media with no audio track and explains it", () => {

src/components/ai-edition/CaptionsPane.tsx

@@ -20,14 +20,18 @@ import {
 } from "@/lib/ai-edition/captions";
 import { useProjectStore } from "@/lib/ai-edition/store/projectStore";
 import {
+	useAssetTranscriptions,
 	useTimelineTranscriptGate,
 	useTranscriptionStore,
 } from "@/lib/ai-edition/store/transcriptionStore";
 import { useCaptions } from "@/lib/ai-edition/store/useCaptions";
+import { firstTimelineBusyView } from "@/lib/ai-edition/transcription/status";
 import { nativeBridgeClient } from "@/native";
 import { ColorField } from "./ColorField";
 import styles from "./NewEditorShell.module.css";
 import { SliderCell, Toggle } from "./RightPanes";
+import { useTranscriptionLabel } from "./TranscriptionStatus";
+import { transcriptionBusyLabel } from "./transcriptionBusyLabel";
 
 /** The families `src/index.css` already loads for on-canvas text — anything else
  *  would render in the preview but fall back to a default in the export canvas. */
@@ -98,7 +102,17 @@ export function CaptionsPane() {
 	// actual footage had speech.
 	const gate = useTimelineTranscriptGate();
 	const requestTimelineTranscripts = useTranscriptionStore((s) => s.requestTimelineTranscripts);
+	const transcriptions = useAssetTranscriptions();
+	const transcriptionLabel = useTranscriptionLabel();
 	const isTranscribing = gate.state === "pending";
+	// Timeline-scoped on purpose: the gate below answers for the timeline's
+	// assets, so the label must too — an off-timeline job must not relabel an
+	// enabled button.
+	const busyLabel = transcriptionBusyLabel(
+		firstTimelineBusyView(document, transcriptions) ??
+			(isTranscribing ? { assetId: "", status: "running", phase: "loading-model" } : undefined),
+		transcriptionLabel,
+	);
 	const silentMedia = gate.state === "blocked" && gate.reason === "no-audio";
 	const engineError = gate.state === "blocked" && gate.reason === "failed" ? gate.message : null;
 
@@ -232,7 +246,7 @@ export function CaptionsPane() {
 							onClick={() => void requestTimelineTranscripts()}
 						>
 							{isTranscribing ? <Loader2 size={14} className="animate-spin" /> : null}
-							{isTranscribing ? t("captions.transcribing") : t("captions.transcribe")}
+							{busyLabel ?? t("captions.transcribe")}
 						</button>
 					</div>
 				) : (
@@ -245,10 +259,13 @@ export function CaptionsPane() {
 					>
 						{/* The cue count is only meaningful while the layer is on — deriving
 						    cues short-circuits when it's off, so a "0 lines" reading there
-						    would say the transcript is empty when it isn't. */}
-						{settings.enabled
-							? t("captions.derivedFromTranscript", { count: cues.length })
-							: t("captions.hiddenHint")}
+						    would say the transcript is empty when it isn't. While a
+						    regeneration is in flight the phase label matters more than the
+						    count of cues about to be replaced. */}
+						{busyLabel ??
+							(settings.enabled
+								? t("captions.derivedFromTranscript", { count: cues.length })
+								: t("captions.hiddenHint"))}
 					</p>
 				)}
 

src/components/ai-edition/NewEditorShell.tsx

@@ -26,6 +26,7 @@ import { useUndoRedoShortcuts } from "@/lib/ai-edition/store/undo";
 import { useSequentialTimelineOps } from "@/lib/ai-edition/store/useSequentialTimelineOps";
 import { useTimeline } from "@/lib/ai-edition/store/useTimeline";
 import { newRegionDurationSec } from "@/lib/ai-edition/timeline/newRegionDuration";
+import { firstTimelineBusyView } from "@/lib/ai-edition/transcription/status";
 import { matchesShortcut } from "@/lib/shortcuts";
 import { nativeBridgeClient } from "@/native";
 import type { AiEditionProjectSummary } from "@/native/contracts";
@@ -159,6 +160,12 @@ export function NewEditorShell() {
 				.map((v) => v.assetId),
 		[transcriptions],
 	);
+	// For the pane-level busy label: timeline-scoped like the gate, so an
+	// off-timeline job cannot relabel controls the gate keeps enabled.
+	const timelineBusyView = useMemo(
+		() => firstTimelineBusyView(document, transcriptions),
+		[document, transcriptions],
+	);
 	const tl = useTimeline();
 	// An undo only puts the restored document back in the store and marks it dirty,
 	// so without this the reverted state never reached disk: close the window and the
@@ -1148,6 +1155,8 @@ export function NewEditorShell() {
 		assets: document?.assets ?? [],
 		trimRanges: document?.timeline?.trimRanges ?? [],
 		busyAssetIds,
+		transcriptions,
+		busyView: timelineBusyView,
 		onSeek: handleSeek,
 		onAddTrimRange: handleAddTrimRange,
 		onRemoveTrimRange: handleRemoveTrimRange,

src/components/ai-edition/RightPanes.tsx

@@ -62,7 +62,10 @@ import {
 import { hasAnyClipWithCamera } from "@/lib/ai-edition/timeline/camera";
 import { formatMs } from "@/lib/ai-edition/timeline/format";
 import { locateVirtualPosition } from "@/lib/ai-edition/timeline/virtual-preview";
-import type { TranscriptGateReason } from "@/lib/ai-edition/transcription/status";
+import {
+	type AssetTranscriptionView,
+	type TranscriptGateReason,
+} from "@/lib/ai-edition/transcription/status";
 import { getAssetPath } from "@/lib/assetPath";
 import { resolveWebcamLayoutPreset, supportsWebcamReactiveZoom } from "@/lib/compositeLayout";
 import { supportsCursorClickEffects } from "@/lib/cursor/cursorCapabilities";
@@ -81,6 +84,8 @@ import {
 	getAspectRatioLabel,
 } from "@/utils/aspectRatioUtils";
 import styles from "./NewEditorShell.module.css";
+import { useTranscriptionLabel } from "./TranscriptionStatus";
+import { transcriptionBusyLabel } from "./transcriptionBusyLabel";
 
 interface PaneProps {
 	title: string;
@@ -641,6 +646,8 @@ export function TranscriptPane({
 	assets,
 	trimRanges,
 	busyAssetIds,
+	transcriptions,
+	busyView,
 	onSeek,
 	onAddTrimRange,
 	onRemoveTrimRange,
@@ -659,6 +666,12 @@ export function TranscriptPane({
 	 *  stream silently swallow Backspace and hover-bin clicks for the whole
 	 *  background pass, with nothing on screen to say why. */
 	busyAssetIds: readonly string[];
+	transcriptions?: Record<string, AssetTranscriptionView>;
+	/** First busy view over the TIMELINE's assets (same scope as `blocked` and
+	 *  `isTranscribing`) — the pane-level label reads this, never the whole
+	 *  `transcriptions` record, so an off-timeline job cannot relabel controls
+	 *  the gate keeps enabled. */
+	busyView?: AssetTranscriptionView;
 	onSeek: (sec: number) => void;
 	onAddTrimRange: (target: TrimTarget, startSec: number, endSec: number, reason: string) => void;
 	onRemoveTrimRange: (trimId: string) => void;
@@ -706,6 +719,12 @@ export function TranscriptPane({
 	// Only silence is a dead end: every other reason (a retryable failure, no
 	// engine, nothing attempted) leaves the button worth pressing.
 	const silentMedia = blocked?.reason === "no-audio";
+	const transcriptionLabel = useTranscriptionLabel();
+	const paneBusyLabel = transcriptionBusyLabel(
+		busyView ??
+			(isTranscribing ? { assetId: "", status: "running", phase: "loading-model" } : undefined),
+		transcriptionLabel,
+	);
 
 	if (clips.length === 0 || !hasAnyTranscript) {
 		return (
@@ -731,7 +750,7 @@ export function TranscriptPane({
 						{clips.length === 0
 							? ts("transcript.noClips")
 							: isTranscribing
-								? ts("transcript.transcribing")
+								? (paneBusyLabel ?? ts("transcript.transcribing"))
 								: silentMedia
 									? ts("transcript.noAudio")
 									: ts("transcript.noTranscript")}
@@ -749,7 +768,7 @@ export function TranscriptPane({
 						// fail on the same missing track every time.
 						disabled={!canTranscribe || isTranscribing || silentMedia}
 					>
-						{isTranscribing ? ts("transcript.transcribing") : ts("transcript.transcribeNow")}
+						{paneBusyLabel ?? ts("transcript.transcribeNow")}
 					</button>
 				</div>
 			</Pane>
@@ -768,6 +787,10 @@ export function TranscriptPane({
 						index={idx}
 						section={section}
 						busy={busyAssetIds.includes(section.clip.assetId)}
+						busyLabel={
+							transcriptionBusyLabel(transcriptions?.[section.clip.assetId], transcriptionLabel) ??
+							undefined
+						}
 						cueWordId={cueWordId}
 						onSeek={onSeek}
 						onAddTrimRange={onAddTrimRange}
@@ -795,6 +818,7 @@ const TranscriptClipBlock = memo(function TranscriptClipBlock({
 	index,
 	section,
 	busy,
+	busyLabel,
 	cueWordId,
 	onSeek,
 	onAddTrimRange,
@@ -803,6 +827,7 @@ const TranscriptClipBlock = memo(function TranscriptClipBlock({
 	index: number;
 	section: ClipSection;
 	busy: boolean;
+	busyLabel?: string;
 	cueWordId: string | null;
 	onSeek: (sec: number) => void;
 	onAddTrimRange: (target: TrimTarget, startSec: number, endSec: number, reason: string) => void;
@@ -1095,7 +1120,7 @@ const TranscriptClipBlock = memo(function TranscriptClipBlock({
 						}}
 					>
 						<Loader2 size={12} className="animate-spin" />
-						{ts("transcript.transcribing")}
+						{busyLabel ?? ts("transcript.transcribing")}
 					</span>
 				) : null}
 			</span>
@@ -1109,7 +1134,7 @@ const TranscriptClipBlock = memo(function TranscriptClipBlock({
 						fontStyle: "italic",
 					}}
 				>
-					{busy ? ts("transcript.transcribing") : ts("transcript.noClipTranscript")}
+					{busy ? (busyLabel ?? ts("transcript.transcribing")) : ts("transcript.noClipTranscript")}
 				</p>
 			) : (
 				<div

src/components/ai-edition/TranscriptPane.gating.test.tsx

@@ -77,8 +77,9 @@ describe("transcript pane gating", () => {
 
 	it("shows the background run in progress instead of an idle button", () => {
 		renderPane({ isTranscribing: true });
-		const button = screen.getByRole("button", { name: "Transcribing…" });
+		const button = screen.getByRole("button", { name: "Starting speech model" });
 		expect(button).toBeDisabled();
+		expect(screen.queryByRole("button", { name: "Transcribing…" })).toBeNull();
 	});
 
 	it("disables the button when the timeline's media have no audio track, and says why", () => {

src/components/ai-edition/TranscriptionStatus.test.tsx

@@ -61,12 +61,25 @@ describe("useTranscriptionLabel telemetry suffix", () => {
 		);
 	});
 
-	// The model download has its own words and no device to report yet.
-	it("leaves the model-download phase alone", () => {
+	// Cold start / already-cached model: loading-model with no byte counters is
+	// the helper booting, not a download, so the copy must not say "Downloading".
+	it("names initializing when loading-model has no in-flight bytes", () => {
 		expect(labelOf(running({ phase: "loading-model", backend: "whispercpp-cpu", rtf: 1.1 }))).toBe(
-			"mediaStage.downloadingModel",
+			"mediaStage.initializingModel",
 		);
 	});
+
+	it("names downloading when loading-model reports an in-flight byte count", () => {
+		expect(
+			labelOf(
+				running({
+					phase: "loading-model",
+					downloadedBytes: 40_000_000,
+					totalBytes: 253_000_000,
+				}),
+			),
+		).toBe("mediaStage.downloadingModel");
+	});
 });
 
 describe("TranscriptionStatusDot", () => {

src/components/ai-edition/TranscriptionStatus.tsx

@@ -12,6 +12,7 @@ import { useScopedT } from "@/contexts/I18nContext";
 import {
 	type AssetTranscriptionView,
 	isCpuBackend,
+	isModelDownloadInFlight,
 	progressFraction,
 	realtimeSpeed,
 } from "@/lib/ai-edition/transcription/status";
@@ -30,7 +31,11 @@ export function useTranscriptionLabel(): (view: AssetTranscriptionView) => strin
 				// screen to explain it, so it gets its own words rather than being
 				// labelled "Transcribing" — this is the phase most often mistaken for
 				// a hang, and the one `phase` was carried through the store for.
-				if (view.phase === "loading-model") return t("mediaStage.downloadingModel");
+				if (view.phase === "loading-model") {
+					return isModelDownloadInFlight(view)
+						? t("mediaStage.downloadingModel")
+						: t("mediaStage.initializingModel");
+				}
 				// Transcribing a long recording runs for minutes. A bare
 				// "Transcribing…" for that whole time is indistinguishable from a
 				// hang, so append the percentage as soon as the main process reports

src/components/ai-edition/transcriptionBusyLabel.test.ts

@@ -0,0 +1,23 @@
+import { describe, expect, it } from "vitest";
+import type { AssetTranscriptionView } from "@/lib/ai-edition/transcription/status";
+import { transcriptionBusyLabel } from "./transcriptionBusyLabel";
+
+function view(extra: Partial<AssetTranscriptionView> = {}): AssetTranscriptionView {
+	return { assetId: "a", status: "running", phase: "loading-model", ...extra };
+}
+
+describe("Captions/Transcript busy copy", () => {
+	it("does not use captions.transcribing while pending on loading-model", () => {
+		const label = transcriptionBusyLabel(view(), (v) =>
+			v.phase === "loading-model" ? "mediaStage.initializingModel" : "captions.transcribing",
+		);
+		expect(label).toBe("mediaStage.initializingModel");
+		expect(label).not.toContain("captions.transcribing");
+	});
+
+	it("returns null when idle so the pane can keep its transcribe verb", () => {
+		expect(
+			transcriptionBusyLabel(view({ status: "idle", phase: undefined }), () => "x"),
+		).toBeNull();
+	});
+});

src/components/ai-edition/transcriptionBusyLabel.ts

@@ -0,0 +1,13 @@
+import type { AssetTranscriptionView } from "@/lib/ai-edition/transcription/status";
+
+/**
+ * Copy for a boolean spinner (Captions / Transcript / Source transcript).
+ * Returns null when nothing is in flight so the caller can keep its idle verb.
+ */
+export function transcriptionBusyLabel(
+	view: AssetTranscriptionView | undefined,
+	labelOf: (view: AssetTranscriptionView) => string,
+): string | null {
+	if (!view || (view.status !== "running" && view.status !== "queued")) return null;
+	return labelOf(view);
+}

src/components/ai-edition/v4/MediaStage.tsx

@@ -447,7 +447,7 @@ export function MediaStage({
 								) : (
 									<span style={{ color: "var(--muted)" }}>
 										{selectedBusy
-											? t("mediaStage.transcribingEllipsis")
+											? transcriptionLabel(selectedTranscription)
 											: selectedTranscription.status === "failed"
 												? t("mediaStage.generationFailedHint")
 												: t("mediaStage.notGeneratedHint")}

src/i18n/locales/ar/editor.json

@@ -145,6 +145,7 @@
 		"notGeneratedHint": "لم يُنشأ بعد — اختر لغة وانقر على إعادة الإنشاء.",
 		"transcribing": "جارٍ النسخ",
 		"downloadingModel": "جارٍ تنزيل نموذج الكلام",
+		"initializingModel": "جارٍ تهيئة نموذج الكلام",
 		"transcribingEllipsis": "جارٍ النسخ…",
 		"pendingTranscription": "بانتظار النسخ",
 		"transcriptionFailed": "فشل النسخ",

src/i18n/locales/ar/settings.json

@@ -275,7 +275,6 @@
 		"show": "إظهار الترجمة",
 		"noTranscript": "تُقرأ الترجمة من نص الوسائط. فرّغ نص هذا الفيديو لتفعيلها.",
 		"transcribe": "تفريغ نص الفيديو",
-		"transcribing": "جارٍ التفريغ…",
 		"derivedFromTranscript": "{{count}} سطر ترجمة، مشتقة مباشرةً من النص المفرّغ.",
 		"hiddenHint": "تأتي الترجمة من النص المفرّغ لهذه الوسائط. فعّلها لرؤيتها في المعاينة وفي التصدير.",
 		"legacyAnnotations": "لا يزال هذا المشروع يحتوي على تعليقات ترجمة من الميزة القديمة ({{count}}). تُرسم فوق طبقة الترجمة.",

src/i18n/locales/en/editor.json

@@ -140,6 +140,7 @@
 		"transcriptReady": "Transcript ready",
 		"transcribing": "Transcribing",
 		"downloadingModel": "Downloading speech model",
+		"initializingModel": "Starting speech model",
 		"transcribingEllipsis": "Transcribing…",
 		"pendingTranscription": "Pending transcription",
 		"transcriptionFailed": "Transcription failed",

src/i18n/locales/en/settings.json

@@ -281,7 +281,6 @@
 		"show": "Show captions",
 		"noTranscript": "Captions are read from the media transcript. Transcribe this video to turn them on.",
 		"transcribe": "Transcribe video",
-		"transcribing": "Transcribing…",
 		"derivedFromTranscript": "{{count}} caption lines, derived live from the transcript.",
 		"hiddenHint": "Captions come from this media's transcript. Turn them on to see them in the preview and in exports.",
 		"legacyAnnotations": "This project still contains caption annotations from the old captions feature ({{count}}). They render on top of the caption layer.",

src/i18n/locales/es/editor.json

@@ -145,6 +145,7 @@
 		"notGeneratedHint": "Aún no generada — elige un idioma y haz clic en regenerar.",
 		"transcribing": "Transcribiendo",
 		"downloadingModel": "Descargando modelo de voz",
+		"initializingModel": "Iniciando el modelo de voz",
 		"transcribingEllipsis": "Transcribiendo…",
 		"pendingTranscription": "Transcripción pendiente",
 		"transcriptionFailed": "Error de transcripción",