@k-l-lambda/lilylet 0.1.76 → 0.1.78

package/source/lilylet/meiEncoder.ts

@@ -419,6 +419,8 @@ const extractMarkOptions = (marks?: Mark[]): {
 	fingerings: { finger: number; placement?: 'above' | 'below' }[];
 	navigation?: 'coda' | 'segno';
 	markups: { content: string; placement?: 'above' | 'below' }[];
+	breath: boolean;
+	glissando: boolean;
 } => {
 	const result = {
 		artics: [] as { type: string; placement?: 'above' | 'below' }[],
@@ -441,6 +443,8 @@ const extractMarkOptions = (marks?: Mark[]): {
 		fingerings: [] as { finger: number; placement?: 'above' | 'below' }[],
 		navigation: undefined as 'coda' | 'segno' | undefined,
 		markups: [] as { content: string; placement?: 'above' | 'below' }[],
+		breath: false,
+		glissando: false,
 	};
 
 	if (!marks) return result;
@@ -472,6 +476,8 @@ const extractMarkOptions = (marks?: Mark[]): {
 					result.mordent = 'lower';
 				} else if (mark.type === OrnamentType.prall) {
 					result.mordent = 'upper';
+				} else if (mark.type === OrnamentType.breath) {
+					result.breath = true;
 				}
 				break;
 			case 'dynamic': {
@@ -537,6 +543,9 @@ const extractMarkOptions = (marks?: Mark[]): {
 					placement: (mark as { placement?: 'above' | 'below' }).placement,
 				});
 				break;
+			case 'glissando':
+				result.glissando = true;
+				break;
 		}
 
 		// Tremolo (special case - from parser internal mark)
@@ -570,6 +579,8 @@ interface NoteEventResult {
 	fingerings: { finger: number; placement?: 'above' | 'below' }[];
 	navigation?: 'coda' | 'segno';
 	markups: { content: string; placement?: 'above' | 'below' }[];
+	breath: boolean;
+	glissando: boolean;
 }
 
 // Convert NoteEvent to MEI
@@ -638,6 +649,8 @@ const noteEventToMEI = (
 			fingerings: markOptions.fingerings,
 			navigation: markOptions.navigation,
 			markups: markOptions.markups,
+			breath: markOptions.breath,
+			glissando: markOptions.glissando,
 		};
 	}
 
@@ -701,6 +714,8 @@ const noteEventToMEI = (
 		fingerings: markOptions.fingerings,
 		navigation: markOptions.navigation,
 		markups: markOptions.markups,
+		breath: markOptions.breath,
+		glissando: markOptions.glissando,
 	};
 };
 
@@ -996,6 +1011,15 @@ interface TrillRef {
 	startid: string;
 }
 
+interface BreathRef {
+	startid: string;   // the note the breath follows
+}
+
+interface GlissRef {
+	startid: string;
+	endid: string;
+}
+
 interface MordentRef {
 	startid: string;
 	form?: 'upper';  // prall = upper mordent
@@ -1044,6 +1068,7 @@ interface SlurSpan {
 // Tie state for cross-measure ties - maps staff:layer to pending pitches
 type TieState = Record<string, Pitch[]>;
 type SlurState = Record<string, string[]>;  // voice key -> open slur startIds (stack, oldest first)
+type GlissState = Record<string, string | null>;  // voice key -> pending glissando start id (awaiting next-measure target)
 type HairpinState = Record<string, { form: 'cres' | 'dim'; startId: string } | null>;  // voice key -> pending hairpin
 
 // Pending octave span for cross-measure continuation
@@ -1069,6 +1094,8 @@ interface LayerResult {
 	arpeggios: ArpegRef[];
 	fermatas: FermataRef[];
 	trills: TrillRef[];
+	breaths: BreathRef[];
+	glissandos: GlissRef[];
 	mordents: MordentRef[];
 	turns: TurnRef[];
 	dynamics: DynamRef[];
@@ -1079,6 +1106,7 @@ interface LayerResult {
 	markups: MarkupRef[];
 	pendingTiePitches: Pitch[];  // For cross-measure tie tracking
 	pendingSlur: string[];  // For cross-measure slur tracking (all open slur startIds)
+	pendingGliss: string | null;  // For cross-measure glissando (start note awaiting its target in the next measure)
 	pendingHairpin: { form: 'cres' | 'dim'; startId: string } | null;  // For cross-measure hairpin tracking
 	pendingOctave: PendingOctave | null;  // For cross-measure ottava span tracking
 	ottavaExplicitlyClosed: boolean;  // True if ottava was closed by explicit \ottava #0 in this layer
@@ -1117,7 +1145,7 @@ const getEventBeamMarks = (event: NoteEvent | RestEvent | TupletEvent | TimesEve
 };
 
 // Encode a layer (voice)
-const encodeLayer = (voice: Voice, layerN: number, indent: string, initialTiePitches: Pitch[] = [], keyFifths: number = 0, initialClef?: Clef, initialSlurs: string[] = [], initialHairpin: { form: 'cres' | 'dim'; startId: string } | null = null, initialOctave: PendingOctave | null = null): LayerResult => {
+const encodeLayer = (voice: Voice, layerN: number, indent: string, initialTiePitches: Pitch[] = [], keyFifths: number = 0, initialClef?: Clef, initialSlurs: string[] = [], initialHairpin: { form: 'cres' | 'dim'; startId: string } | null = null, initialOctave: PendingOctave | null = null, initialGliss: string | null = null): LayerResult => {
 	const layerId = generateId("layer");
 	let xml = `${indent}<layer xml:id="${layerId}" n="${layerN}">\n`;
 
@@ -1197,6 +1225,12 @@ const encodeLayer = (voice: Voice, layerN: number, indent: string, initialTiePit
 	// Track ornament refs
 	const fermatas: FermataRef[] = [];
 	const trills: TrillRef[] = [];
+	const breaths: BreathRef[] = [];
+	// Glissando: a note-to-note slide. Record the start note id and close it on the
+	// NEXT encoded note (endid), mirroring slur start→end pairing. A trailing
+	// glissando with no following note in the layer is dropped (no target).
+	const glissandos: GlissRef[] = [];
+	let pendingGlissStart: string | null = initialGliss;
 	const mordents: MordentRef[] = [];
 	const turns: TurnRef[] = [];
 	const dynamics: DynamRef[] = [];
@@ -1374,6 +1408,19 @@ const encodeLayer = (voice: Voice, layerN: number, indent: string, initialTiePit
 				if (result.trill) {
 					trills.push({ startid: result.elementId });
 				}
+				// Breath mark follows this note.
+				if (result.breath) {
+					breaths.push({ startid: result.elementId });
+				}
+				// Glissando: close a pending start onto this note, then open a new one
+				// if this note itself carries a glissando.
+				if (pendingGlissStart) {
+					glissandos.push({ startid: pendingGlissStart, endid: result.elementId });
+					pendingGlissStart = null;
+				}
+				if (result.glissando) {
+					pendingGlissStart = result.elementId;
+				}
 				if (result.mordent) {
 					mordents.push({
 						startid: result.elementId,
@@ -1645,7 +1692,7 @@ const encodeLayer = (voice: Voice, layerN: number, indent: string, initialTiePit
 	}
 
 	xml += `${indent}</layer>\n`;
-	return { xml, hairpins, pedals, octaves, slurs, arpeggios, fermatas, trills, mordents, turns, dynamics, fingerings, navigations, harmonies, barlines, markups, pendingTiePitches, pendingSlur: openSlurs.map(s => s.startId), pendingHairpin, pendingOctave, ottavaExplicitlyClosed, endingClef: currentClef, lastNoteId, currentOttavaShift, octaveEndReplacements };
+	return { xml, hairpins, pedals, octaves, slurs, arpeggios, fermatas, trills, breaths, glissandos, mordents, turns, dynamics, fingerings, navigations, harmonies, barlines, markups, pendingTiePitches, pendingSlur: openSlurs.map(s => s.startId), pendingGliss: pendingGlissStart, pendingHairpin, pendingOctave, ottavaExplicitlyClosed, endingClef: currentClef, lastNoteId, currentOttavaShift, octaveEndReplacements };
 };
 
 // Staff result type
@@ -1658,6 +1705,8 @@ interface StaffResult {
 	arpeggios: ArpegRef[];
 	fermatas: FermataRef[];
 	trills: TrillRef[];
+	breaths: BreathRef[];
+	glissandos: GlissRef[];
 	mordents: MordentRef[];
 	turns: TurnRef[];
 	dynamics: DynamRef[];
@@ -1668,6 +1717,7 @@ interface StaffResult {
 	markups: MarkupRef[];
 	pendingTies: TieState;  // For cross-measure tie tracking
 	pendingSlurs: SlurState;  // For cross-measure slur tracking
+	pendingGlisses: GlissState;  // For cross-measure glissando tracking
 	pendingHairpins: HairpinState;  // For cross-measure hairpin tracking
 	pendingOctaves: OttavaState;  // For cross-measure ottava span tracking
 	ottavaExplicitlyClosed: Record<string, boolean>;  // Track which layers had ottava explicitly closed
@@ -1677,7 +1727,7 @@ interface StaffResult {
 }
 
 // Encode a staff
-const encodeStaff = (voices: Voice[], staffN: number, indent: string, tieState: TieState = {}, slurState: SlurState = {}, hairpinState: HairpinState = {}, ottavaState: OttavaState = {}, keyFifths: number = 0, initialClef?: Clef): StaffResult => {
+const encodeStaff = (voices: Voice[], staffN: number, indent: string, tieState: TieState = {}, slurState: SlurState = {}, hairpinState: HairpinState = {}, ottavaState: OttavaState = {}, keyFifths: number = 0, initialClef?: Clef, glissState: GlissState = {}): StaffResult => {
 	const staffId = generateId("staff");
 	let xml = `${indent}<staff xml:id="${staffId}" n="${staffN}">\n`;
 	const allHairpins: HairpinSpan[] = [];
@@ -1687,6 +1737,8 @@ const encodeStaff = (voices: Voice[], staffN: number, indent: string, tieState:
 	const allArpeggios: ArpegRef[] = [];
 	const allFermatas: FermataRef[] = [];
 	const allTrills: TrillRef[] = [];
+	const allBreaths: BreathRef[] = [];
+	const allGlissandos: GlissRef[] = [];
 	const allMordents: MordentRef[] = [];
 	const allTurns: TurnRef[] = [];
 	const allDynamics: DynamRef[] = [];
@@ -1697,6 +1749,7 @@ const encodeStaff = (voices: Voice[], staffN: number, indent: string, tieState:
 	const allMarkups: MarkupRef[] = [];
 	const pendingTies: TieState = {};
 	const pendingSlurs: SlurState = {};
+	const pendingGlisses: GlissState = {};
 	const pendingHairpins: HairpinState = {};
 	const pendingOctaves: OttavaState = {};
 	const ottavaExplicitlyClosed: Record<string, boolean> = {};
@@ -1714,7 +1767,8 @@ const encodeStaff = (voices: Voice[], staffN: number, indent: string, tieState:
 			const initialSlur = slurState[tieKey] || [];
 			const initialHairpin = hairpinState[tieKey] || null;
 			const initialOctave = ottavaState[tieKey] || null;
-			const result = encodeLayer(voice, layerN, indent + '    ', initialTies, keyFifths, endingClef, initialSlur, initialHairpin, initialOctave);
+			const initialGliss = glissState[tieKey] || null;
+			const result = encodeLayer(voice, layerN, indent + '    ', initialTies, keyFifths, endingClef, initialSlur, initialHairpin, initialOctave, initialGliss);
 			xml += result.xml;
 			allHairpins.push(...result.hairpins);
 			allPedals.push(...result.pedals);
@@ -1723,6 +1777,8 @@ const encodeStaff = (voices: Voice[], staffN: number, indent: string, tieState:
 			allArpeggios.push(...result.arpeggios);
 			allFermatas.push(...result.fermatas);
 			allTrills.push(...result.trills);
+			allBreaths.push(...result.breaths);
+			allGlissandos.push(...result.glissandos);
 			allMordents.push(...result.mordents);
 			allTurns.push(...result.turns);
 			allDynamics.push(...result.dynamics);
@@ -1740,6 +1796,7 @@ const encodeStaff = (voices: Voice[], staffN: number, indent: string, tieState:
 			// Always record (even empty) so a measure that closed all its slurs
 			// clears the carried stack rather than leaving a stale open slur.
 			pendingSlurs[tieKey] = result.pendingSlur;
+			pendingGlisses[tieKey] = result.pendingGliss;
 			// Track pending hairpins for this layer
 			if (result.pendingHairpin) {
 				pendingHairpins[tieKey] = result.pendingHairpin;
@@ -1771,6 +1828,8 @@ const encodeStaff = (voices: Voice[], staffN: number, indent: string, tieState:
 		arpeggios: allArpeggios,
 		fermatas: allFermatas,
 		trills: allTrills,
+		breaths: allBreaths,
+		glissandos: allGlissandos,
 		mordents: allMordents,
 		turns: allTurns,
 		dynamics: allDynamics,
@@ -1781,6 +1840,7 @@ const encodeStaff = (voices: Voice[], staffN: number, indent: string, tieState:
 		markups: allMarkups,
 		pendingTies,
 		pendingSlurs,
+		pendingGlisses,
 		pendingHairpins,
 		pendingOctaves,
 		ottavaExplicitlyClosed,
@@ -1903,7 +1963,7 @@ const BARLINE_TO_MEI: Record<string, string> = {
 
 // Encode a measure
 // encodeMeasure accepts mutable tieState, slurState, hairpinState, ottavaState and clefState that persist across measures
-const encodeMeasure = (measure: Measure, measureN: number, indent: string, totalStaves: number, tieState: TieState, slurState: SlurState, hairpinState: HairpinState, ottavaState: OttavaState, keyFifths: number = 0, partInfos: PartInfo[] = [], clefState: ClefState = {}, octaveEndReplacements: Record<string, string> = {}, measureIds?: string[]): string => {
+const encodeMeasure = (measure: Measure, measureN: number, indent: string, totalStaves: number, tieState: TieState, slurState: SlurState, hairpinState: HairpinState, ottavaState: OttavaState, keyFifths: number = 0, partInfos: PartInfo[] = [], clefState: ClefState = {}, octaveEndReplacements: Record<string, string> = {}, measureIds?: string[], glissState: GlissState = {}): string => {
 	const measureId = generateId("measure");
 	if (measureIds) measureIds.push(measureId);
 	let staffContent = '';  // Build staff content first, then add measure tag with barline
@@ -1914,6 +1974,8 @@ const encodeMeasure = (measure: Measure, measureN: number, indent: string, total
 	const allArpeggios: ArpegRef[] = [];
 	const allFermatas: FermataRef[] = [];
 	const allTrills: TrillRef[] = [];
+	const allBreaths: BreathRef[] = [];
+	const allGlissandos: GlissRef[] = [];
 	const allMordents: MordentRef[] = [];
 	const allTurns: TurnRef[] = [];
 	const allDynamics: DynamRef[] = [];
@@ -1963,7 +2025,7 @@ const encodeMeasure = (measure: Measure, measureN: number, indent: string, total
 	for (let si = 1; si <= totalStaves; si++) {
 		const voices = voicesByStaff[si] || [];
 		const initialClef = clefState[si];
-		const result = encodeStaff(voices, si, indent + '    ', tieState, slurState, hairpinState, ottavaState, keyFifths, initialClef);
+		const result = encodeStaff(voices, si, indent + '    ', tieState, slurState, hairpinState, ottavaState, keyFifths, initialClef, glissState);
 		staffContent += result.xml;
 		allHairpins.push(...result.hairpins);
 		allPedals.push(...result.pedals);
@@ -1972,6 +2034,8 @@ const encodeMeasure = (measure: Measure, measureN: number, indent: string, total
 		allArpeggios.push(...result.arpeggios);
 		allFermatas.push(...result.fermatas);
 		allTrills.push(...result.trills);
+		allBreaths.push(...result.breaths);
+		allGlissandos.push(...result.glissandos);
 		allMordents.push(...result.mordents);
 		allTurns.push(...result.turns);
 		allDynamics.push(...result.dynamics);
@@ -1985,6 +2049,7 @@ const encodeMeasure = (measure: Measure, measureN: number, indent: string, total
 		Object.assign(tieState, result.pendingTies);
 		// Update slur state with pending slurs from this staff
 		Object.assign(slurState, result.pendingSlurs);
+		Object.assign(glissState, result.pendingGlisses);
 		// Update hairpin state with pending hairpins from this staff
 		Object.assign(hairpinState, result.pendingHairpins);
 		// Update ottava state with pending octaves from this staff.
@@ -2066,6 +2131,16 @@ const encodeMeasure = (measure: Measure, measureN: number, indent: string, total
 		staffContent += `${indent}    <trill xml:id="${generateId('trill')}" startid="#${tr.startid}" />\n`;
 	}
 
+	// Generate breath marks (the breath follows its startid note)
+	for (const br of allBreaths) {
+		staffContent += `${indent}    <breath xml:id="${generateId('breath')}" startid="#${br.startid}" />\n`;
+	}
+
+	// Generate glissando spans (note-to-note slide)
+	for (const gl of allGlissandos) {
+		staffContent += `${indent}    <gliss xml:id="${generateId('gliss')}" startid="#${gl.startid}" endid="#${gl.endid}" />\n`;
+	}
+
 	// Generate mordent control events
 	for (const mord of allMordents) {
 		const formAttr = mord.form ? ` form="${mord.form}"` : '';
@@ -2812,6 +2887,7 @@ const encode = (doc: LilyletDoc, options: MEIEncoderOptions = {}): string => {
 
 	// Track slur state across measures for cross-measure slurs
 	const slurState: SlurState = {};
+	const glissState: GlissState = {};
 
 	// Track hairpin state across measures for cross-measure hairpins
 	const hairpinState: HairpinState = {};
@@ -2901,7 +2977,7 @@ const encode = (doc: LilyletDoc, options: MEIEncoderOptions = {}): string => {
 		const openTag = segOpenAt.get(mIndex);
 		if (openTag) mei += openTag;
 		const measureIndent = segDecomp ? `${mIndent}${indent}` : mIndent;
-		mei += encodeMeasure(measure, mIndex, measureIndent, totalStaves, tieState, slurState, hairpinState, ottavaState, currentKey, partInfos, clefState, octaveEndReplacements, measureIds);
+		mei += encodeMeasure(measure, mIndex, measureIndent, totalStaves, tieState, slurState, hairpinState, ottavaState, currentKey, partInfos, clefState, octaveEndReplacements, measureIds, glissState);
 		const closeTag = segCloseAt.get(mIndex);
 		if (closeTag) mei += closeTag;
 	});

package/source/lilylet/musicXmlDecoder.ts

@@ -25,6 +25,7 @@ import {
 	HairpinType,
 	PedalType,
 	NavigationMarkType,
+	OrnamentType,
 	Placement,
 	BarlineEvent,
 	HarmonyEvent,
@@ -305,12 +306,13 @@ class VoiceTracker {
 			});
 			this.currentStaff.set(voiceNum, staff);
 		}
-		const voice = this.voices.get(voiceNum)!;
-		// Update staff if specified
-		if (staff > 0) {
-			voice.staff = staff;
-		}
-		return voice;
+		// NOTE: do NOT overwrite voice.staff on later notes. voice.staff is the voice's
+		// HOME staff — the staff of its first note — matching the parser, which fixes
+		// voice.staff to the first leading \staff (so `\staff "1" … \staff "2" …` gives
+		// voice.staff=1, with the switch carried by an inline ctx(staff=2) event). The
+		// previous code clobbered it to every note's staff, so a voice that ended on a
+		// different staff round-tripped with the wrong home staff.
+		return this.voices.get(voiceNum)!;
 	}
 
 	addEvent(voiceNum: number, event: Event, duration: number, staff: number = 1): void {
@@ -430,9 +432,11 @@ const parseNotations = (notationsEl: Element): MusicXmlNotations => {
 		}
 	}
 
-	// Fermata
-	if (hasElement(notationsEl, 'fermata')) {
-		result.fermata = true;
+	// Fermata. The "angled" shape is LilyPond's \shortfermata; everything else
+	// (normal/square/unset) is the ordinary \fermata.
+	const fermataEl = getElements(notationsEl, 'fermata')[0];
+	if (fermataEl) {
+		result.fermata = (fermataEl.textContent || '').trim() === 'angled' ? 'short' : 'normal';
 	}
 
 	// Arpeggiate
@@ -457,6 +461,14 @@ const parseNotations = (notationsEl: Element): MusicXmlNotations => {
 		};
 	}
 
+	// Glissando / slide — both are note-to-note slide lines. Only the START note
+	// carries the mark (LilyPond \glissando auto-connects to the next note), so we
+	// record presence of any start; stops are ignored.
+	const slideEls = [...getElements(notationsEl, 'glissando'), ...getElements(notationsEl, 'slide')];
+	if (slideEls.some(el => getAttribute(el, 'type') === 'start')) {
+		result.glissando = true;
+	}
+
 	return result;
 };
 
@@ -726,6 +738,17 @@ const parseDirection = (dirEl: Element): MusicXmlDirection => {
 		result.segno = true;
 	}
 
+	// Rehearsal mark (<direction-type><rehearsal>A</rehearsal>). May be split into
+	// multiple <rehearsal> (e.g. "A" + a Chinese gloss "段"); join them. Collected
+	// from the whole direction so it is found regardless of which direction-type it
+	// sits in. Surfaced as text → markup by directionToMarks (per the text-as-markup
+	// convention; LilyPond's musicxml2ly emits \mark \markup { \box { … } }).
+	const rehearsalEls = Array.from(dirEl.getElementsByTagName('rehearsal'));
+	if (rehearsalEls.length > 0) {
+		const text = rehearsalEls.map(el => el.textContent?.trim() || '').filter(Boolean).join(' ');
+		if (text) result.rehearsal = text;
+	}
+
 	return result;
 };
 
@@ -811,6 +834,21 @@ const parseMetadata = (doc: Document): Metadata => {
 		metadata.title = movementTitleEl.textContent.trim();
 	}
 
+	// Subtitle: MusicXML has no <subtitle> element — it lives in a page <credit>
+	// tagged <credit-type>subtitle</credit-type>, with the text in <credit-words>.
+	const creditEls = doc.getElementsByTagName('credit');
+	for (let i = 0; i < creditEls.length; i++) {
+		const credit = creditEls[i];
+		const typeEl = credit.getElementsByTagName('credit-type')[0];
+		if (typeEl?.textContent?.trim().toLowerCase() === 'subtitle') {
+			const words = Array.from(credit.getElementsByTagName('credit-words'))
+				.map(w => w.textContent?.trim() || '')
+				.filter(Boolean)
+				.join(' ');
+			if (words) { metadata.subtitle = words; break; }
+		}
+	}
+
 	// Identification (composer, arranger, lyricist)
 	const identificationEl = doc.getElementsByTagName('identification')[0];
 	if (identificationEl) {
@@ -890,6 +928,12 @@ const notationsToMarks = (
 	// Articulations
 	if (notations.articulations) {
 		for (const artName of notations.articulations) {
+			// breath-mark is a NoDirection articulation in MusicXML but maps to
+			// LilyPond's \breathe — model it as the breath ornament.
+			if (artName === 'breath-mark') {
+				marks.push({ markType: 'ornament', type: OrnamentType.breath });
+				continue;
+			}
 			const artType = convertArticulation(artName);
 			if (artType) {
 				marks.push({ markType: 'articulation', type: artType });
@@ -909,7 +953,7 @@ const notationsToMarks = (
 
 	// Fermata
 	if (notations.fermata) {
-		marks.push({ markType: 'ornament', type: 'fermata' as any });
+		marks.push({ markType: 'ornament', type: (notations.fermata === 'short' ? 'shortFermata' : 'fermata') as any });
 	}
 
 	// Arpeggiate
@@ -917,6 +961,11 @@ const notationsToMarks = (
 		marks.push({ markType: 'ornament', type: 'arpeggio' as any });
 	}
 
+	// Glissando / slide (start note only)
+	if (notations.glissando) {
+		marks.push({ markType: 'glissando' });
+	}
+
 	return marks;
 };
 
@@ -1082,16 +1131,31 @@ const directionToMarks = (
 		marks.push({ markType: 'navigation', type: NavigationMarkType.segno });
 	}
 
+	// Rehearsal mark → markup (text-as-markup convention; LilyPond uses \mark).
+	if (direction.rehearsal) {
+		const placement = direction.placement === 'above' ? Placement.above
+			: direction.placement === 'below' ? Placement.below
+			: undefined;
+		marks.push({ markType: 'markup', content: direction.rehearsal, placement });
+	}
+
 	// Words (text directions: "dolce", "espr.", "cresc.", "con forza", ...).
 	// Tempo words ("Allegro", "a tempo", ...) are consumed separately as a tempo
-	// ContextChange by directionToContextChange, so skip those here to avoid
-	// double-emitting; everything else becomes a markup mark → MEI <dir>. Metronome
-	// directions are tempo too, never markup. Navigation text ("D.C. al Fine",
-	// "To Coda", "Fine", "D.S. al Coda") is NOT a tempo word, so it flows here as a
-	// markup glyph — the jump SEMANTICS are captured separately as measure-layout.
+	// ContextChange by directionToContextChange (which takes words[0]), so skip that
+	// one here to avoid double-emitting; the REST become a markup mark → MEI <dir>.
+	// A single <direction> often carries a tempo word AND an expressive/translation
+	// word together (e.g. <words>Moderato</words><words>中板</words>) — only the tempo
+	// word is consumed; the rest must still surface as markup. Metronome directions
+	// are tempo too, never markup. Navigation text ("D.C. al Fine", "To Coda",
+	// "Fine", "D.S. al Coda") is NOT a tempo word, so it flows here as a markup glyph
+	// — the jump SEMANTICS are captured separately as measure-layout.
 	let emittedWords = false;
 	if (direction.words && direction.words.length > 0 && !direction.metronome) {
-		const text = direction.words.map(w => w.text).join('').trim();
+		// If words[0] was consumed as the tempo text, drop it and keep the remainder;
+		// otherwise the whole run is one expressive phrase, joined as today.
+		const firstIsTempo = isTempoWord(direction.words[0].text.trim());
+		const rest = firstIsTempo ? direction.words.slice(1) : direction.words;
+		const text = rest.map(w => w.text).join('').trim();
 		if (text && !isTempoWord(text)) {
 			const placement = direction.placement === 'above' ? Placement.above
 				: direction.placement === 'below' ? Placement.below
@@ -1338,7 +1402,16 @@ const convertMeasure = (
 			// hairpin stop that lands just before a rest (common in piano scores:
 			// `<pedal stop/>` followed by rests filling the voice) is silently dropped.
 			const marks: Mark[] = note.isRest ? [] : (pendingMarks.get(voiceNum) || []);
-			if (!note.isRest) pendingMarks.delete(voiceNum);
+			if (!note.isRest) {
+				pendingMarks.delete(voiceNum);
+				// Drain directions that named this note's staff but had no known voice
+				// yet (queued under the -staff placeholder key by the <direction> handler).
+				const staffPending = pendingMarks.get(-staffNum);
+				if (staffPending && staffPending.length > 0) {
+					marks.push(...staffPending);
+					pendingMarks.delete(-staffNum);
+				}
+			}
 
 			if (note.isRest) {
 				// Rest event
@@ -1366,7 +1439,7 @@ const convertMeasure = (
 				// so the encoder can emit <fermata startid="#rest">; without this the
 				// 3-of-4 fermatas that sit on rests in typical piano scores are lost.
 				if (note.notations?.fermata) {
-					restEvent.marks = [{ markType: 'ornament', type: 'fermata' as any }];
+					restEvent.marks = [{ markType: 'ornament', type: (note.notations.fermata === 'short' ? 'shortFermata' : 'fermata') as any }];
 				}
 
 				// Grace notes don't advance time
@@ -1480,9 +1553,17 @@ const convertMeasure = (
 			// Handle marks (dynamics, hairpins, etc.)
 			const marks = directionToMarks(direction, spannerTracker);
 			if (marks.length > 0) {
-				// Store marks to attach to next note in current voice
-				const existing = pendingMarks.get(currentVoice) || [];
-				pendingMarks.set(currentVoice, [...existing, ...marks]);
+				// Route to the voice the marks belong to. A <direction> in a grand-staff
+				// part names its <staff>; the matching note is the NEXT note on that staff
+				// — NOT currentVoice, which after a <backup> still points at the voice that
+				// just ended (so a staff-2 direction emitted after the backup would wrongly
+				// land on staff 1). Queue under a -staff placeholder key; the next note on
+				// that staff drains it, and the end-of-measure flush falls back to that
+				// staff's last note if no further note appears. Without a staff (single-
+				// staff parts), fall back to currentVoice as before.
+				const targetKey = direction.staff ? -direction.staff : currentVoice;
+				const existing = pendingMarks.get(targetKey) || [];
+				pendingMarks.set(targetKey, [...existing, ...marks]);
 			}
 		} else if (tagName === 'backup') {
 			// A <forward> with no note after it (before this backup) is cursor
@@ -1560,10 +1641,23 @@ const convertMeasure = (
 			}
 			return undefined;
 		};
+		// A staff-keyed leftover (negative key = -staffNum, queued for a "next note on
+		// this staff" that never came): resolve to the last note on any voice that lives
+		// on that staff.
+		const findLastNoteOnStaff = (staff: number): NoteEvent | undefined => {
+			let best: NoteEvent | undefined;
+			for (const [, vs] of allVoices) {
+				if ((vs.staff || 1) !== staff) continue;
+				for (let i = vs.events.length - 1; i >= 0; i--) {
+					if (vs.events[i].type === 'note') { best = vs.events[i] as NoteEvent; break; }
+				}
+			}
+			return best;
+		};
 		for (const [voiceNum, marks] of pendingMarks) {
 			if (marks.length === 0) continue;
-			let target = findLastNote(voiceNum);
-			// Voice had no note (e.g. direction-only or rest-only): attach to the
+			let target = voiceNum < 0 ? findLastNoteOnStaff(-voiceNum) : findLastNote(voiceNum);
+			// Voice/staff had no note (e.g. direction-only or rest-only): attach to the
 			// last note of any voice so the marking is not silently dropped.
 			if (!target) {
 				for (const vn of allVoices.keys()) {