fix(core,engine): make a group's preview match what the render bakes

Findings 8-11 and 16 — the render/preview parity set. All group audio,
all invisible to preview or to export alone, which is why each had a
passing test beside it.

9 — The preview bus wired only the automation lane, and
readElementAutomation reads data-automation, so a group's own
`data-volume` never reached the graph at all: the bus stayed at unity
while the render applied it. The export was ~8 dB quieter than what had
been auditioned. The test named for this asserted only
`resolves.not.toBeNull()`.

10 — The volume lane was scheduled on the FX chain's SOURCE, ahead of
the effects, breaking the contract scheduleVolumeLane's own docstring
states: "the volume lane rides the fader, after the effects — where a
DAW puts it, and the order the render bakes it in". The element path
honoured it; the group path did not, so any nonlinear group effect
previewed differently than it rendered. Both now land on a dedicated
post-FX fader node: input → chain → fader → mute → output.

11 — The bus deliberately outlives stopAll(), and groupInput early
-returned on an existing entry, so after a replay or a seek no new ramps
were booked and the gain held the previous pass's last value — 0 after a
fade-out, i.e. silent for the rest of the session. It re-anchors once
per play generation now (ElementFxHandle gains `reanchor`), and the
record keeps its `fx` handle so setRate can re-aim a group's automation,
which its docblock already claimed it did.

8 — The sub-mix summed members at unity (normalize=0) into a pcm_s16le
intermediate, so an over-unity sum hard-clipped at ±1 BEFORE the group's
FX and fader ran: pulling the fader down, or the Giant preset's
compressor, then operated on distortion. The intermediate is float now.
Both downstream readers already took float; applyVolumeEnvelopeToWav did
not, and does now, so a group's envelope is still baked sample-accurately
rather than silently degrading to the expression path.

The new level test is the one that matters here: EVERY tone in that file
is built on ffmpeg's `sine` source, which peaks at ~0.125 full scale, so
nothing in the suite could reach a clip whatever gain it asked for.
`writePeakTone` states the amplitude outright, and the test fails against
the 16-bit intermediate.

16 — mixGroupMembers forked mixAudioTracks' filter build and dropped its
automation-degradation retry, so a member envelope past this ffmpeg
build's expression limits failed the whole composition's audio where an
ungrouped one degrades to base volume with a warning. The retry is back,
reported on a successful result the same way. The group track is also
pushed with its volumeKeyframes when the envelope could not be baked,
instead of losing the group's automation silently.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

packages/core/src/audioGroups.ts

@@ -35,7 +35,13 @@ export interface HfAudioGroup {
   hidden: boolean;
 }
 
-function parseGroupVolume(el: Element | undefined): number {
+/**
+ * A group element's `data-volume`, defaulting to 1 for a missing, unparseable
+ * or absent element. Shared so the preview bus and `resolveAudioGroups` (which
+ * the render reads) cannot drift: the export was ~8 dB quieter than what was
+ * auditioned for exactly as long as the preview ignored this.
+ */
+export function readAudioGroupVolume(el: Element | null | undefined): number {
   const raw = el?.getAttribute("data-volume");
   const parsed = raw ? parseFloat(raw) : 1;
   return Number.isFinite(parsed) ? parsed : 1;
@@ -50,7 +56,7 @@ function buildGroup(id: string, memberIds: string[], el: Element | undefined): H
     memberIds,
     ...(fxChain ? { fxChain } : {}),
     ...(automation ? { automation } : {}),
-    volume: parseGroupVolume(el),
+    volume: readAudioGroupVolume(el),
     hidden: el?.hasAttribute("data-hidden") ?? false,
   };
 }

packages/core/src/runtime/audioFx.ts

@@ -98,6 +98,16 @@ export function readElementAutomation(el: {
  */
 export interface ElementFxHandle {
   dispose(): void;
+  /**
+   * Re-book every envelope against a fresh reference frame.
+   *
+   * For a graph that OUTLIVES the source feeding it — a group bus, which is
+   * built once and kept for the session — a replay or a seek starts a new pass
+   * over the same chain. Without re-anchoring, the envelopes stay committed to
+   * the first pass's absolute context times: past their last point that is a
+   * stuck value, which for a fade-out is silence for the rest of the session.
+   */
+  reanchor(timing: AutomationTiming): void;
   /**
    * Re-aim every booked envelope at a new playback rate.
    *
@@ -277,6 +287,13 @@ export function attachElementFxChain(
   }
 
   return {
+    reanchor: (next: AutomationTiming) => {
+      if (disposed) return;
+      const at = timingNow();
+      if (at) cancelParamLane(automated, at.scheduledAt);
+      frame = { ...next };
+      scheduleFor(readChain(el).chain, frame);
+    },
     setRate: (rate: number) => {
       const at = timingNow();
       if (disposed || !at || !Number.isFinite(rate) || rate <= 0 || rate === at.rate) return;

packages/core/src/runtime/webAudioTransport.test.ts

@@ -726,16 +726,17 @@ describe("WebAudioTransport", () => {
       await scheduleGrouped(transport, gen, "b", "vo");
 
       // Creation order for a: a-gain(0), a-solo(1), groupInput(2), groupOutput(3),
-      // muteGain(4) — the group bus is built lazily inside a's schedule call.
-      // Then b: b-gain(5), b-solo(6).
-      expect(mock.gainNodes.length).toBeGreaterThanOrEqual(7);
+      // muteGain(4), fader(5) — the group bus is built lazily inside a's
+      // schedule call. Then b: b-gain(6), b-solo(7).
+      expect(mock.gainNodes.length).toBeGreaterThanOrEqual(8);
       const aGain = mock.gainNodes[0]!;
       const aSolo = mock.gainNodes[1]!;
       const groupInput = firstGroupInput(mock);
       const groupOutput = mock.gainNodes[3]!;
       const muteGain = mock.gainNodes[4]!;
-      const bGain = mock.gainNodes[5]!;
-      const bSolo = mock.gainNodes[6]!;
+      const fader = mock.gainNodes[5]!;
+      const bGain = mock.gainNodes[6]!;
+      const bSolo = mock.gainNodes[7]!;
 
       // Each member feeds its own solo gain, and both solo gains feed the
       // shared bus — neither connects straight to master.
@@ -746,11 +747,14 @@ describe("WebAudioTransport", () => {
       expect(aSolo.connect).not.toHaveBeenCalledWith(mock.masterGain);
       expect(bSolo.connect).not.toHaveBeenCalledWith(mock.masterGain);
 
-      // The bus's input never reaches master directly — it lands on the mute
-      // gain (B5) first (the dry passthrough, since neither member's group has
-      // a chain-bearing `<hf-audio-group>`), then the output gain, then master.
+      // The bus's input never reaches master directly. It runs through the
+      // chain (dry here — neither member's group has a chain-bearing
+      // `<hf-audio-group>`) onto the FADER, then the mute gain (B5), then the
+      // output gain, then master. The fader sits POST-FX because that is where
+      // the render bakes group volume in.
       expect(groupInput.connect).not.toHaveBeenCalledWith(mock.masterGain);
-      expect(groupInput.connect).toHaveBeenCalledWith(muteGain);
+      expect(groupInput.connect).toHaveBeenCalledWith(fader);
+      expect(fader.connect).toHaveBeenCalledWith(muteGain);
       expect(muteGain.connect).toHaveBeenCalledWith(groupOutput);
       expect(groupOutput.connect).toHaveBeenCalledWith(mock.masterGain);
     });
@@ -773,18 +777,36 @@ describe("WebAudioTransport", () => {
 
       const muteGain = mock.gainNodes[4]!;
       const groupOutput = mock.gainNodes[3]!;
-      expect(firstGroupInput(mock).connect).toHaveBeenCalledWith(muteGain);
+      const fader = mock.gainNodes[5]!;
+      expect(firstGroupInput(mock).connect).toHaveBeenCalledWith(fader);
+      expect(fader.connect).toHaveBeenCalledWith(muteGain);
       expect(muteGain.connect).toHaveBeenCalledWith(groupOutput);
       expect(groupOutput.connect).toHaveBeenCalledWith(mock.masterGain);
+      // No element to read, so the fader sits at unity.
+      expect(fader.gain.value).toBe(1);
     });
 
-    it("group volume rides the group's own data-volume via its automation lane, not the member's", async () => {
+    // The old assertion here was `resolves.not.toBeNull()` — it was named for
+    // group volume and checked only that scheduling did not throw, so the bus
+    // sitting at unity while the render applied data-volume went unseen. The
+    // export was ~8 dB quieter than what had been auditioned.
+    it("puts the group's own data-volume on the bus fader", async () => {
+      document.body.innerHTML = `<hf-audio-group id="vo" data-label="Voiceover" data-volume="0.4"></hf-audio-group>`;
+      const { transport, mock, gen } = setupGroupTransport();
+
+      await expect(scheduleGrouped(transport, gen, "a", "vo")).resolves.not.toBeNull();
+
+      expect(mock.gainNodes[5]!.gain.value).toBeCloseTo(0.4, 6);
+    });
+
+    it("leaves the fader at unity when the group carries no data-volume", async () => {
       document.body.innerHTML = `<hf-audio-group id="vo" data-label="Voiceover"></hf-audio-group>`;
-      const { transport, gen } = setupGroupTransport();
+      const { transport, mock, gen } = setupGroupTransport();
 
       // No throw wiring the group's automation reader against a real
       // <hf-audio-group> element that carries no fx/automation attrs.
       await expect(scheduleGrouped(transport, gen, "a", "vo")).resolves.not.toBeNull();
+      expect(mock.gainNodes[5]!.gain.value).toBe(1);
     });
 
     it("destroy() disposes every group bus", async () => {
@@ -811,13 +833,37 @@ describe("WebAudioTransport", () => {
       expect(mock.gainNodes.filter((n) => n === groupInput)).toHaveLength(1);
     });
 
+    // Surviving stopAll() is the point of the bus — and the trap. Its envelopes
+    // were booked against the FIRST pass's absolute context times, so a replay
+    // or a seek left the fader holding that pass's last value: 0 after a
+    // fade-out, i.e. silent for the rest of the session.
+    it("re-anchors the reused bus once per play generation, and only once", async () => {
+      document.body.innerHTML = `<hf-audio-group id="vo" data-volume="0.5"></hf-audio-group>`;
+      const { transport, mock, gen } = setupGroupTransport();
+      await scheduleGrouped(transport, gen, "a", "vo");
+      const fader = mock.gainNodes[5]!;
+
+      // Something moved the fader mid-pass (a ramp reaching its last point).
+      fader.gain.value = 0;
+      transport.stopAll();
+
+      const gen2 = transport.startGeneration();
+      await scheduleGrouped(transport, gen2, "a", "vo");
+      expect(fader.gain.value).toBeCloseTo(0.5, 6);
+
+      // A second member in the SAME pass must not re-book on top of the first.
+      fader.gain.value = 0;
+      await scheduleGrouped(transport, gen2, "b", "vo");
+      expect(fader.gain.value).toBe(0);
+    });
+
     describe('solo — "Hear only this" (B5)', () => {
       it("silences a non-soloed member via its own solo gain, without touching the group bus", async () => {
         const { transport, mock, gen } = setupGroupTransport();
         await scheduleGrouped(transport, gen, "a", "vo");
         await scheduleGrouped(transport, gen, "b", "vo");
         const aSolo = mock.gainNodes[1]!;
-        const bSolo = mock.gainNodes[6]!;
+        const bSolo = mock.gainNodes[7]!;
         const groupInput = firstGroupInput(mock);
 
         transport.setSolo(new Set(["other-clip"]));
@@ -834,7 +880,7 @@ describe("WebAudioTransport", () => {
         await scheduleGrouped(transport, gen, "a", "vo");
         await scheduleGrouped(transport, gen, "b", "vo");
         const aSolo = mock.gainNodes[1]!;
-        const bSolo = mock.gainNodes[6]!;
+        const bSolo = mock.gainNodes[7]!;
         const groupInput = firstGroupInput(mock);
 
         transport.setSolo(new Set(["a"]));
@@ -849,7 +895,7 @@ describe("WebAudioTransport", () => {
         await scheduleGrouped(transport, gen, "a", "vo");
         await scheduleGrouped(transport, gen, "b", "vo");
         const aSolo = mock.gainNodes[1]!;
-        const bSolo = mock.gainNodes[6]!;
+        const bSolo = mock.gainNodes[7]!;
 
         transport.setSolo(new Set(["vo"]));
 

packages/core/src/runtime/webAudioTransport.ts

@@ -5,7 +5,7 @@ import {
   type AutomationTiming,
 } from "../audio/audioFxAutomation.js";
 import { VOLUME_RANGE } from "../audioAutomation.js";
-import { audioGroupOf, isAudibleUnderSolo } from "../audioGroups.js";
+import { audioGroupOf, isAudibleUnderSolo, readAudioGroupVolume } from "../audioGroups.js";
 import { swallow } from "./diagnostics";
 import { clampAudioGain } from "../audioGain.js";
 import { getDebugSurface } from "./globals.js";
@@ -136,9 +136,17 @@ export class WebAudioTransport {
     string,
     {
       input: GainNode;
+      /** Post-FX fader: `data-volume` plus the volume lane. */
+      fader: GainNode;
       muteGain: GainNode;
       analyser: AnalyserNode;
       levelBuf: Float32Array;
+      /** Kept so `setRate` can re-aim this bus's FX automation, the way it does
+       *  every source's — its docblock claims it already did. */
+      fx: ElementFxHandle | null;
+      /** Play generation the current envelopes were booked against. */
+      generation: number;
+      reanchor(timing: AutomationTiming): void;
       dispose(): void;
     }
   >();
@@ -308,7 +316,18 @@ export class WebAudioTransport {
    */
   private groupInput(groupId: string, doc: Document, timing: AutomationTiming): GainNode | null {
     const existing = this._groups.get(groupId);
-    if (existing) return existing.input;
+    if (existing) {
+      // The bus outlives `stopAll()` on purpose, so a replay or a seek reuses
+      // this graph — but its envelopes were committed to the FIRST pass's
+      // absolute context times. Left alone they hold their last value forever,
+      // which for a fade-out is silence for the rest of the session. Re-anchor
+      // once per play generation, not once per member scheduled.
+      if (existing.generation !== this._playGeneration) {
+        existing.generation = this._playGeneration;
+        existing.reanchor(timing);
+      }
+      return existing.input;
+    }
     if (!this._ctx || !this._masterGain) return null;
 
     const input = this._ctx.createGain();
@@ -329,24 +348,42 @@ export class WebAudioTransport {
     const muteGain = this._ctx.createGain();
     muteGain.gain.value = groupEl?.hasAttribute("data-hidden") ? 0 : 1;
     muteGain.connect(output);
+    // The group's fader, POST-FX: `data-volume` is the static position and the
+    // volume lane rides it, which is where a DAW puts it and the order the
+    // render bakes it in (`scheduleVolumeLane`'s own contract). Scheduling it
+    // on `input` instead put the fader ahead of the effects, so any nonlinear
+    // group effect — a compressor, the Giant preset — previewed differently
+    // than it rendered.
+    const fader = this._ctx.createGain();
+    fader.gain.value = readAudioGroupVolume(groupEl);
+    fader.connect(muteGain);
     const fx = attachElementFxChain(
       this._ctx,
       groupEl ?? { getAttribute: () => null },
       input,
-      muteGain,
+      fader,
       timing,
     );
-    if (groupEl) scheduleVolumeLane(groupEl, input, timing);
+    if (groupEl) scheduleVolumeLane(groupEl, fader, timing);
 
     this._groups.set(groupId, {
       input,
+      fader,
       muteGain,
       analyser,
       levelBuf: new Float32Array(analyser.fftSize),
+      fx,
+      generation: this._playGeneration,
+      reanchor: (at: AutomationTiming) => {
+        fader.gain.value = readAudioGroupVolume(groupEl);
+        fx?.reanchor(at);
+        if (groupEl) scheduleVolumeLane(groupEl, fader, at);
+      },
       dispose: () => {
         try {
           fx?.dispose();
           input.disconnect();
+          fader.disconnect();
           muteGain.disconnect();
           output.disconnect();
           analyser.disconnect();
@@ -588,6 +625,16 @@ export class WebAudioTransport {
         swallow("webAudioTransport.setRate", err);
       }
     }
+    // Group buses are not in `_activeSources` — they outlive it — so their FX
+    // automation needs re-aiming here too, or a rate change leaves a group's
+    // envelopes running the old plan over audio at the new speed.
+    for (const group of this._groups.values()) {
+      try {
+        group.fx?.setRate(safeRate);
+      } catch (err) {
+        swallow("webAudioTransport.setRate.group", err);
+      }
+    }
     return true;
   }
 

packages/engine/src/services/audioMixer.grouping.test.ts

@@ -63,6 +63,35 @@ function writeTone(path: string, freq: number, seconds: number, gain: number): v
   if (result.status !== 0) throw new Error(`Could not write tone: ${result.stderr}`);
 }
 
+/**
+ * A sine of an EXACT peak amplitude.
+ *
+ * `writeTone` builds on ffmpeg's `sine` source, whose output is ~0.125 full
+ * scale, so its `gain` argument is a relative knob rather than a level: a tone
+ * asked for at 0.7 lands at about -21 dBFS. Fine for the relative comparisons
+ * above, useless for anything about headroom — `aevalsrc` states the amplitude
+ * outright.
+ */
+function writePeakTone(path: string, freq: number, seconds: number, peak: number): void {
+  const result = spawnSync(
+    getFfmpegBinary(),
+    [
+      "-nostdin",
+      "-v",
+      "error",
+      "-f",
+      "lavfi",
+      "-i",
+      `aevalsrc=${peak}*sin(2*PI*${freq}*t):d=${seconds}:s=48000`,
+      "-c:a",
+      "pcm_s16le",
+      path,
+    ],
+    { encoding: "utf-8" },
+  );
+  if (result.status !== 0) throw new Error(`Could not write tone: ${result.stderr}`);
+}
+
 const track = (id: string, end: number, volume = 1) => ({
   id,
   src: `${id}.wav`,
@@ -268,4 +297,54 @@ describe.skipIf(!HAS_FFMPEG)("mix level arithmetic", () => {
     const flatTail = meanVolumeDb(flatOut, 3, 4);
     expect(Math.abs(groupedTail - flatTail)).toBeLessThan(0.5);
   });
+
+  // Members sum at unity (normalize=0), so an over-unity sum used to hard-clip
+  // at ±1 in the 16-bit intermediate BEFORE the group's fader and FX chain ran
+  // — pulling the group down then operated on distortion. Every existing probe
+  // here sums to ≤ 0.8, which is exactly why nothing caught it; preview cannot
+  // reproduce it either, because its bus is float.
+  it("does not clip an over-unity member sum before the group fader", async () => {
+    const projectDir = mkdtempSync(join(tmpdir(), "hf-grp-clip-"));
+    const workDir = mkdtempSync(join(tmpdir(), "hf-grp-clip-work-"));
+    tempDirs.push(projectDir, workDir);
+
+    // Two coherent copies of the same tone: 0.7 + 0.7 = 1.4, comfortably over.
+    // `writePeakTone`, not `writeTone` — ffmpeg's `sine` source is nowhere near
+    // full scale (its output peaks at ~0.125), so every tone in this file sits
+    // around -21 dBFS and NOTHING here can reach a clip no matter what gain is
+    // asked for. That is a large part of why this class of bug survived.
+    writePeakTone(join(projectDir, "a.wav"), 440, 2, 0.7);
+    writePeakTone(join(projectDir, "b.wav"), 440, 2, 0.7);
+    // The reference: the level that sum SHOULD reach once the group's 0.5
+    // fader has been applied — 1.4 × 0.5 = 0.7, one tone's worth.
+    writePeakTone(join(projectDir, "ref.wav"), 440, 2, 0.7);
+
+    const groupedOut = join(projectDir, `clip-grouped-${MIXED_AUDIO_FILENAME}`);
+    const refOut = join(projectDir, `clip-ref-${MIXED_AUDIO_FILENAME}`);
+
+    const grouped = await processCompositionAudio(
+      [
+        { ...track("a", 2), groupId: "vo", groupVolume: 0.5 },
+        { ...track("b", 2), groupId: "vo", groupVolume: 0.5 },
+      ],
+      projectDir,
+      workDir,
+      groupedOut,
+      2,
+    );
+    const reference = await processCompositionAudio(
+      [track("ref", 2)],
+      projectDir,
+      workDir,
+      refOut,
+      2,
+    );
+    expect(grouped.success).toBe(true);
+    expect(reference.success).toBe(true);
+
+    // Both are ONE track into the outer mix, so the outer graph is identical
+    // and the levels are directly comparable. Clipped, the flat-topped sum
+    // reads well over a dB hot even after the fader halves it.
+    expect(Math.abs(meanVolumeDb(groupedOut) - meanVolumeDb(refOut))).toBeLessThan(0.5);
+  });
 });

packages/engine/src/services/audioMixer.ts

@@ -885,20 +885,25 @@ async function mixGroupMembers(
   totalDuration: number,
   signal?: AbortSignal,
   config?: Partial<Pick<EngineConfig, "ffmpegProcessTimeout">>,
-): Promise<{ success: boolean; error?: string }> {
+): Promise<{ success: boolean; error?: string; degradedAutomation?: boolean }> {
   const ffmpegProcessTimeout = config?.ffmpegProcessTimeout ?? DEFAULT_CONFIG.ffmpegProcessTimeout;
   const outputDir = dirname(outputPath);
   if (!existsSync(outputDir)) mkdirSync(outputDir, { recursive: true });
 
-  const inputFilters = memberTracks.map((track, i) => {
-    const delayMs = Math.round(track.start * 1000);
-    const trimDuration = track.end - track.start + (track.tailSeconds ?? 0);
-    const volumeFilter = buildVolumeExpression(track);
-    return `[${i}:a]atrim=0:${formatFilterNumber(trimDuration)},${volumeFilter},adelay=${delayMs}|${delayMs},apad,atrim=0:${formatFilterNumber(totalDuration)}[a${i}]`;
-  });
+  const buildInputFilters = (ignoreKeyframes: boolean) =>
+    memberTracks.map((track, i) => {
+      const delayMs = Math.round(track.start * 1000);
+      const trimDuration = track.end - track.start + (track.tailSeconds ?? 0);
+      const volumeFilter = buildVolumeExpression(track, ignoreKeyframes);
+      return `[${i}:a]atrim=0:${formatFilterNumber(trimDuration)},${volumeFilter},adelay=${delayMs}|${delayMs},apad,atrim=0:${formatFilterNumber(totalDuration)}[a${i}]`;
+    });
   const mixInputs = memberTracks.map((_, i) => `[a${i}]`).join("");
 
-  const runOnce = async (useNormalize: boolean): Promise<RunFfmpegResult> => {
+  const runOnce = async (
+    useNormalize: boolean,
+    ignoreKeyframes = false,
+  ): Promise<RunFfmpegResult> => {
+    const inputFilters = buildInputFilters(ignoreKeyframes);
     const mixFilter = useNormalize
       ? `${mixInputs}amix=inputs=${memberTracks.length}:duration=longest:dropout_transition=0:normalize=0[out]`
       : // amix's default normalize divides by input count; compensate by THIS
@@ -922,8 +927,15 @@ async function mixGroupMembers(
         scriptPath,
         "-map",
         "[out]",
+        // Float, not pcm_s16le: `normalize=0` sums the members at unity, so any
+        // over-unity sum hard-clipped at ±1 in the intermediate — BEFORE the
+        // group's FX chain and its fader ran. Pulling the group down, or the
+        // Giant preset's compressor, then operated on distortion. Preview
+        // cannot reproduce it (its bus is float), and it only shows up in the
+        // export. Both readers downstream take float: `readWav` (format 3) and
+        // `applyVolumeEnvelopeToWav`.
         "-acodec",
-        "pcm_s16le",
+        "pcm_f32le",
         "-ar",
         "48000",
         "-t",
@@ -943,14 +955,32 @@ async function mixGroupMembers(
     }
   };
 
-  let result = await runOnce(true);
+  let useNormalize = true;
+  let result = await runOnce(useNormalize);
   if (!result.success && groupNormalizeOptionUnsupported(result.stderr)) {
-    result = await runOnce(false);
+    useNormalize = false;
+    result = await runOnce(useNormalize);
   }
+
+  // The same defence `mixAudioTracks` has, which this forked without: a
+  // member's volume automation becomes an ffmpeg `volume` expression whose
+  // evaluator limits are build-dependent, so a dense envelope can fail the
+  // whole run. Ungrouped, that track degrades to base volume with a warning;
+  // grouped, it took the entire composition's audio down with it.
+  let degradedAutomation = false;
+  const hasAutomation = memberTracks.some((track) => (track.volumeKeyframes?.length ?? 0) > 0);
+  if (!result.success && !signal?.aborted && hasAutomation) {
+    const retry = await runOnce(useNormalize, true);
+    if (retry.success) {
+      result = retry;
+      degradedAutomation = true;
+    }
+  }
+
   if (signal?.aborted) return { success: false, error: "Group sub-mix cancelled" };
   if (!result.success)
     return { success: false, error: formatFfmpegError(result.exitCode, result.stderr) };
-  return { success: true };
+  return { success: true, degradedAutomation };
 }
 
 export async function processCompositionAudio(
@@ -1305,6 +1335,7 @@ export async function processCompositionAudio(
   // clock is composition time (offset 0): a group has no `data-start` of its
   // own, and members are already delayed to their composition positions
   // inside the sub-mix, so the group WAV's t=0 IS composition time.
+  const groupsDegradedAutomation: string[] = [];
   for (const [groupId, memberTracks] of groupTracks) {
     const meta = groupMeta.get(groupId);
     if (!meta) continue;
@@ -1329,6 +1360,7 @@ export async function processCompositionAudio(
       });
       continue;
     }
+    if (subMix.degradedAutomation) groupsDegradedAutomation.push(groupId);
 
     // Composition-time automation (offset 0, duration totalDuration) — same
     // resolve/lane/bake path a member uses, just anchored at the group clock
@@ -1385,6 +1417,10 @@ export async function processCompositionAudio(
       mediaStart: 0,
       duration: totalDuration,
       volume: bakedEnvelope ? 1.0 : meta.volume,
+      // Same fallback an ungrouped track gets: when the envelope could not be
+      // baked into the samples, hand the keyframes to the outer mix's volume
+      // expression instead of dropping the group's automation on the floor.
+      ...(bakedEnvelope || !laneKeyframes?.length ? {} : { volumeKeyframes: laneKeyframes }),
     });
   }
   if (failures.length > 0) return bail();
@@ -1397,9 +1433,18 @@ export async function processCompositionAudio(
     /* ignore */
   }
 
+  // A group whose sub-mix had to drop member automation reports it the same
+  // way mixAudioTracks reports its own degradation: on a SUCCESSFUL result, so
+  // the render ships and the caller can still say what was lost.
+  const degradedGroups = [...groupsDegradedAutomation];
+  const degradedNote =
+    degradedGroups.length > 0
+      ? `Volume automation exceeded this ffmpeg build's expression limits in group(s) ${degradedGroups.join(", ")}; rendered at base volume`
+      : undefined;
+
   return {
     ...mixResult,
     durationMs: Date.now() - startMs,
-    error: mixResult.error,
+    error: mixResult.error ?? degradedNote,
   };
 }

packages/engine/src/services/audioVolumeEnvelope.test.ts

@@ -173,4 +173,76 @@ describe("applyVolumeEnvelopeToWav", () => {
       ),
     ).toBe(false);
   });
+
+  // The group sub-mix writes float precisely so an over-unity member sum keeps
+  // its headroom until the group's own fader and FX act on it. If the baker
+  // could not read that file it returned false, and the group's automation was
+  // dropped on the floor by the caller.
+  describe("32-bit float input", () => {
+    /** Stereo float32 WAV, every sample `value` — over 1.0 on purpose. */
+    function writeConstantFloatWav(path: string, frames: number, value: number): void {
+      const dataSize = frames * CHANNELS * 4;
+      const buffer = Buffer.alloc(44 + dataSize);
+      buffer.write("RIFF", 0, "ascii");
+      buffer.writeUInt32LE(36 + dataSize, 4);
+      buffer.write("WAVE", 8, "ascii");
+      buffer.write("fmt ", 12, "ascii");
+      buffer.writeUInt32LE(16, 16);
+      buffer.writeUInt16LE(3, 20); // WAVE_FORMAT_IEEE_FLOAT
+      buffer.writeUInt16LE(CHANNELS, 22);
+      buffer.writeUInt32LE(SAMPLE_RATE, 24);
+      buffer.writeUInt32LE(SAMPLE_RATE * CHANNELS * 4, 28);
+      buffer.writeUInt16LE(CHANNELS * 4, 32);
+      buffer.writeUInt16LE(32, 34);
+      buffer.write("data", 36, "ascii");
+      buffer.writeUInt32LE(dataSize, 40);
+      for (let i = 0; i < frames * CHANNELS; i += 1) buffer.writeFloatLE(value, 44 + i * 4);
+      writeFileSync(path, buffer);
+    }
+
+    const floatSampleAt = (path: string, frame: number, channel = 0): number =>
+      readFileSync(path).readFloatLE(44 + (frame * CHANNELS + channel) * 4);
+
+    it("scales float samples and keeps them above 1.0 unclamped", () => {
+      const path = join(tmp(), "float.wav");
+      writeConstantFloatWav(path, SAMPLE_RATE, 1.4);
+
+      expect(
+        applyVolumeEnvelopeToWav(
+          path,
+          [
+            { time: 0, volume: 1 },
+            { time: 1, volume: 1 },
+          ],
+          0,
+          1,
+        ),
+      ).toBe(true);
+
+      // Unity envelope: unchanged, and NOT clamped down to 1.0.
+      expect(floatSampleAt(path, 0)).toBeCloseTo(1.4, 5);
+      expect(floatSampleAt(path, SAMPLE_RATE - 1)).toBeCloseTo(1.4, 5);
+    });
+
+    it("applies the envelope across the file", () => {
+      const path = join(tmp(), "float-fade.wav");
+      writeConstantFloatWav(path, SAMPLE_RATE, 1.4);
+
+      expect(
+        applyVolumeEnvelopeToWav(
+          path,
+          [
+            { time: 0, volume: 1 },
+            { time: 1, volume: 0 },
+          ],
+          0,
+          1,
+        ),
+      ).toBe(true);
+
+      expect(floatSampleAt(path, 0)).toBeCloseTo(1.4, 5);
+      expect(floatSampleAt(path, Math.floor(SAMPLE_RATE / 2))).toBeCloseTo(0.7, 2);
+      expect(floatSampleAt(path, SAMPLE_RATE - 1)).toBeCloseTo(0, 3);
+    });
+  });
 });

packages/engine/src/services/audioVolumeEnvelope.ts

@@ -22,13 +22,17 @@ import { normaliseEnvelope } from "@hyperframes/core/media-volume-envelope";
 import { riffChunks } from "./wavChunks.js";
 
 const PCM_FORMAT = 1; // WAVE_FORMAT_PCM
-const SUPPORTED_BITS = 16;
+const FLOAT_FORMAT = 3; // WAVE_FORMAT_IEEE_FLOAT
 
 interface WavLayout {
   numChannels: number;
   sampleRate: number;
   dataOffset: number;
   dataSize: number;
+  /** 16-bit integer, or 32-bit float — the group sub-mix writes float so an
+   *  over-unity member sum is not hard-clipped before the group's own FX and
+   *  fader get to act on it. */
+  float: boolean;
 }
 
 /**
@@ -39,34 +43,49 @@ interface WavLayout {
  * and trailing chunks (LIST/fact/etc.) are skipped harmlessly. Returns null on
  * anything unexpected so the caller falls back to the expression path.
  */
-function parseWavLayout(buffer: Buffer): WavLayout | null {
-  if (buffer.length < 12 || buffer.toString("ascii", 0, 4) !== "RIFF") return null;
-  if (buffer.toString("ascii", 8, 12) !== "WAVE") return null;
+interface WavFmt {
+  numChannels: number;
+  sampleRate: number;
+  /** 16-bit integer PCM, or 32-bit IEEE float. Anything else is unreadable. */
+  float: boolean;
+}
 
-  let fmt: { numChannels: number; sampleRate: number; bitsPerSample: number } | null = null;
+/** The `fmt ` chunk, or null for a format this cannot safely edit in place. */
+function readFmtChunk(buffer: Buffer, body: number): WavFmt | null {
+  const format = buffer.readUInt16LE(body);
+  const bits = buffer.readUInt16LE(body + 14);
+  const float = format === FLOAT_FORMAT;
+  if (!float && format !== PCM_FORMAT) return null;
+  if (bits !== (float ? 32 : 16)) return null;
+  const numChannels = buffer.readUInt16LE(body + 2);
+  if (numChannels < 1) return null;
+  return { numChannels, sampleRate: buffer.readUInt32LE(body + 4), float };
+}
+
+function isRiffWave(buffer: Buffer): boolean {
+  return (
+    buffer.length >= 12 &&
+    buffer.toString("ascii", 0, 4) === "RIFF" &&
+    buffer.toString("ascii", 8, 12) === "WAVE"
+  );
+}
+
+function parseWavLayout(buffer: Buffer): WavLayout | null {
+  if (!isRiffWave(buffer)) return null;
+
+  let fmt: WavFmt | null = null;
   let data: { offset: number; size: number } | null = null;
 
   for (const { id, body, size } of riffChunks(buffer)) {
     if (id === "fmt " && body + 16 <= buffer.length) {
-      if (buffer.readUInt16LE(body) !== PCM_FORMAT) return null;
-      fmt = {
-        numChannels: buffer.readUInt16LE(body + 2),
-        sampleRate: buffer.readUInt32LE(body + 4),
-        bitsPerSample: buffer.readUInt16LE(body + 14),
-      };
+      fmt = readFmtChunk(buffer, body);
     } else if (id === "data") {
       data = { offset: body, size: Math.min(size, buffer.length - body) };
     }
   }
 
   if (!fmt || !data) return null;
-  if (fmt.bitsPerSample !== SUPPORTED_BITS || fmt.numChannels < 1) return null;
-  return {
-    numChannels: fmt.numChannels,
-    sampleRate: fmt.sampleRate,
-    dataOffset: data.offset,
-    dataSize: data.size,
-  };
+  return { ...fmt, dataOffset: data.offset, dataSize: data.size };
 }
 
 /**
@@ -103,11 +122,41 @@ export function createEnvelopeWalker(
   };
 }
 
+/**
+/** Every sample scaled by the envelope, in place, in whichever of the two
+ *  formats the layout reports. Float is NOT clamped: it is the format the group
+ *  sub-mix writes precisely so an over-unity sum keeps its headroom until
+ *  something downstream chooses to reduce it. */
+function scaleSamples(
+  buffer: Buffer,
+  layout: WavLayout,
+  gainAt: (seconds: number) => number,
+): void {
+  const { numChannels, sampleRate, dataOffset, dataSize, float } = layout;
+  const bytesPerSample = float ? 4 : 2;
+  const frameBytes = numChannels * bytesPerSample;
+  const frameCount = Math.floor(dataSize / frameBytes);
+  const scaleOne = float
+    ? (at: number, gain: number) => buffer.writeFloatLE(buffer.readFloatLE(at) * gain, at)
+    : (at: number, gain: number) => {
+        const scaled = Math.round(buffer.readInt16LE(at) * gain);
+        buffer.writeInt16LE(scaled < -32768 ? -32768 : scaled > 32767 ? 32767 : scaled, at);
+      };
+
+  for (let frame = 0; frame < frameCount; frame += 1) {
+    const gain = gainAt(frame / sampleRate);
+    const base = dataOffset + frame * frameBytes;
+    for (let channel = 0; channel < numChannels; channel += 1) {
+      scaleOne(base + channel * bytesPerSample, gain);
+    }
+  }
+}
+
 /**
  * Multiply a prepared WAV's samples by a time-varying gain envelope in place.
  *
- * @returns `true` if the envelope was applied; `false` if the file isn't the
- *   expected 16-bit PCM (caller should fall back to the expression path).
+ * @returns `true` if the envelope was applied; `false` if the file is neither
+ *   16-bit PCM nor 32-bit float (caller should fall back to the expression path).
  */
 export function applyVolumeEnvelopeToWav(
   wavPath: string,
@@ -123,20 +172,7 @@ export function applyVolumeEnvelopeToWav(
     const layout = parseWavLayout(buffer);
     if (!layout) return false;
 
-    const { numChannels, sampleRate, dataOffset, dataSize } = layout;
-    const bytesPerSample = SUPPORTED_BITS / 8;
-    const frameBytes = numChannels * bytesPerSample;
-    const frameCount = Math.floor(dataSize / frameBytes);
-
-    for (let frame = 0; frame < frameCount; frame += 1) {
-      const gain = gainAt(frame / sampleRate);
-      const base = dataOffset + frame * frameBytes;
-      for (let channel = 0; channel < numChannels; channel += 1) {
-        const at = base + channel * bytesPerSample;
-        const scaled = Math.round(buffer.readInt16LE(at) * gain);
-        buffer.writeInt16LE(scaled < -32768 ? -32768 : scaled > 32767 ? 32767 : scaled, at);
-      }
-    }
+    scaleSamples(buffer, layout, gainAt);
 
     // Write to a uniquely-named sibling then atomically rename over the
     // original. The random name avoids following a pre-planted symlink at a