@koda-sl/baker-cli 0.112.1 → 0.113.1

package/dist/cli.js

@@ -2,14 +2,17 @@
 import {
   ELEVENLABS_MAX_MUSIC_LENGTH_MS,
   IMAGE_GENERATE_MODELS,
+  LayerExecutionError,
   MODEL_REGISTRY,
   SEEDANCE_DURATIONS,
   ValidationError,
   createEngineFromEnv,
   defaultRegistry,
+  describeFailureReason,
   generateCatalog,
+  resolveConcurrency,
   validateCanvasDeep
-} from "./chunk-3JVYU72O.js";
+} from "./chunk-7K2YAWUT.js";
 
 // src/cli.ts
 import { defineCommand as defineCommand155, runMain } from "citty";
@@ -11126,6 +11129,10 @@ var runCommand = defineCommand83({
     "outputs-dir": { type: "string", description: "Per-run outputs root (default ./canvas)" },
     "run-id": { type: "string", description: "Override run id" },
     "cache-policy": { type: "string", description: "read_write | bypass | read_only" },
+    concurrency: {
+      type: "string",
+      description: "Max nodes per layer in flight at once (default 5; env BAKER_CANVAS_CONCURRENCY)"
+    },
     "keep-runs": {
       type: "string",
       description: "After the run, prune old r_* run dirs, keeping the N newest (off by default)"
@@ -11173,7 +11180,11 @@ var runCommand = defineCommand83({
       const policy = args["cache-policy"] ?? "read_write";
       const result = await engine.run(parsed, {
         run_id: args["run-id"] ? String(args["run-id"]) : void 0,
-        cache_policy: policy
+        cache_policy: policy,
+        concurrency: resolveConcurrency(
+          args.concurrency !== void 0 ? String(args.concurrency) : void 0,
+          process.env.BAKER_CANVAS_CONCURRENCY
+        )
       });
       const keepRuns = args["keep-runs"] !== void 0 ? Number(args["keep-runs"]) : void 0;
       if (keepRuns !== void 0 && Number.isFinite(keepRuns)) {
@@ -11203,6 +11214,14 @@ var runCommand = defineCommand83({
         );
         process.exit(2);
       }
+      if (e instanceof LayerExecutionError) {
+        const failures = e.failures.map((f) => ({ node_id: f.nodeId, message: describeFailureReason(f.reason) }));
+        process.stderr.write(
+          `${JSON.stringify({ ok: false, error: { code: "runtime", message: e.message, failures } }, null, 2)}
+`
+        );
+        process.exit(1);
+      }
       const msg = e instanceof Error ? e.message : String(e);
       process.stderr.write(`${JSON.stringify({ ok: false, error: { code: "runtime", message: msg } }, null, 2)}
 `);
@@ -11534,8 +11553,9 @@ function buildDescribeCanvas(imageSource, imageIsUrl, describeModel, selectModel
           prompt: LAYOUT_PROMPT
         }
       }
-    ],
-    output: { node: "select", output: "text" }
+    ]
+    // No declared output: the engine's prune-to-output would drop `layout` (nothing
+    // consumes it) and runVisionPasses reads all THREE node outputs after the run.
   };
 }
 async function runVisionPasses(canvas) {
@@ -11896,7 +11916,43 @@ var NARRATOR_SPEAKERS = /* @__PURE__ */ new Set([
   "offscreen",
   "off-screen"
 ]);
-var SHARED_ASPECT_RATIOS = /* @__PURE__ */ new Set(["1:1", "16:9", "9:16", "4:3", "3:4", "21:9"]);
+var SEEDANCE_GEN_ASPECTS = ["1:1", "3:4", "9:16", "4:3", "16:9", "21:9"];
+var OUTPUT_ASPECTS = ["1:1", "16:9", "9:16", "4:3", "3:4", "21:9", "4:5", "2:3", "3:2"];
+function parseRatio(ar) {
+  const m = ar?.trim().match(/^(\d+(?:\.\d+)?)[:x/](\d+(?:\.\d+)?)$/);
+  if (!m) return void 0;
+  const w = Number(m[1]);
+  const h = Number(m[2]);
+  return w > 0 && h > 0 ? w / h : void 0;
+}
+function nearestAspect(ratio, candidates) {
+  let best = candidates[0];
+  let bestDelta = Number.POSITIVE_INFINITY;
+  for (const c of candidates) {
+    const r = parseRatio(c);
+    if (r === void 0) continue;
+    const delta = Math.abs(ratio - r);
+    if (delta < bestDelta) {
+      best = c;
+      bestDelta = delta;
+    }
+  }
+  return best;
+}
+function resolveAspect(sourceAr, override, genAspects = SEEDANCE_GEN_ASPECTS) {
+  let outAr;
+  if (override !== void 0) {
+    if (!OUTPUT_ASPECTS.includes(override)) {
+      throw new Error(`unsupported --aspect "${override}" \u2014 supported aspect ratios: ${OUTPUT_ASPECTS.join(", ")}`);
+    }
+    outAr = override;
+  } else {
+    const ratio = parseRatio(sourceAr);
+    outAr = ratio === void 0 ? "9:16" : nearestAspect(ratio, OUTPUT_ASPECTS);
+  }
+  const genAr = genAspects.includes(outAr) ? outAr : nearestAspect(parseRatio(outAr), genAspects);
+  return { outAr, genAr, remapped: genAr !== outAr };
+}
 var EDGES = ["start", "end"];
 function snapToSeedance(durationS) {
   if (!Number.isFinite(durationS) || durationS <= 0) return SEEDANCE_DURATIONS[0];
@@ -11934,6 +11990,12 @@ function canvasDims(ar) {
       return { w: 1080, h: 1440 };
     case "21:9":
       return { w: 1920, h: 822 };
+    case "4:5":
+      return { w: 1080, h: 1350 };
+    case "2:3":
+      return { w: 1080, h: 1620 };
+    case "3:2":
+      return { w: 1620, h: 1080 };
     default:
       return { w: 1080, h: 1920 };
   }
@@ -11996,13 +12058,14 @@ function stillHoldArgs(durationS, dims) {
     "{{out.video}}"
   ];
 }
-function trimArgs(durationS, offsetS = 0) {
+function trimArgs(durationS, offsetS = 0, dims) {
   return [
     "-i",
     "{{in.clip}}",
     ...offsetS > 0 ? ["-ss", offsetS.toFixed(3)] : [],
     "-t",
     durationS.toFixed(3),
+    ...dims ? ["-vf", `scale=${dims.w}:${dims.h}:force_original_aspect_ratio=increase,crop=${dims.w}:${dims.h},setsar=1`] : [],
     "-an",
     "-c:v",
     "libx264",
@@ -12207,9 +12270,16 @@ function presenceOf(el) {
   }
   return map;
 }
-function aspectRatioParam(blueprint) {
-  const ar = blueprint.source?.aspect_ratio;
-  return ar && SHARED_ASPECT_RATIOS.has(ar) ? ar : void 0;
+function aspectRemapTodo(aspect) {
+  if (!aspect.remapped) return {};
+  const dims = canvasDims(aspect.outAr);
+  const trims = (parseRatio(aspect.genAr) ?? 0) > (parseRatio(aspect.outAr) ?? 0) ? "sides" : "top/bottom";
+  return {
+    aspect_remap: `Output is ${aspect.outAr} (${dims.w}\xD7${dims.h}) but the video model generates ${aspect.genAr}; every clip is center-cropped onto the ${aspect.outAr} canvas before the spine. Keep subjects centered in frame prompts \u2014 the crop trims the ${trims}.`
+  };
+}
+function genAspectsFor(videoModel) {
+  return /\bveo\b|\/veo-/.test(videoModel) ? ["16:9", "9:16"] : SEEDANCE_GEN_ASPECTS;
 }
 function annotateBlueprintWithElements(blueprintInput, elementsInput) {
   if (!blueprintInput || typeof blueprintInput !== "object") return blueprintInput;
@@ -12350,7 +12420,10 @@ function buildElementSheets(slots, nodes) {
       inputs: { references: [slot.ref] },
       params: {
         model: ACTOR_SHEET_MODEL,
-        subject_description: slot.description ?? `the ${slot.type}`,
+        // The clean-plate clause mirrors the frame prompts' CLEAN PLATE block: a sheet
+        // that comes back with a fake camera app baked in (P2-22) poisons EVERY frame
+        // grounded on it, so the suppression must live on the sheet too.
+        subject_description: `${slot.description ?? `the ${slot.type}`} \u2014 clean plate: no phone-camera UI chrome, no app interface, no watermarks, no captions or on-image text`,
         subject_type: subjectType,
         // 4K: the sheet packs up to 8 cells (angles + tight face/detail close-ups), and
         // it's the ONE reference every frame grounds on — per-cell sharpness here
@@ -12453,6 +12526,10 @@ function buildFramePrompt(edge, sceneIndex, framePrompt, present, hasAnchor, mod
     "{{target_blueprint}}"
   ].join("\n");
 }
+function silentClipTranscript(scene, native) {
+  if (native) return "";
+  return (scene.transcript_slice ?? []).map((w) => w.text?.trim()).filter(Boolean).join(" ").trim();
+}
 function ingestFrameRef(url, edge, ctx, nodes) {
   const cached2 = ctx.ingestCache?.get(url);
   if (cached2) return cached2;
@@ -12479,7 +12556,7 @@ function buildFrameRef(edge, url, framePrompt, present, ctx, nodes) {
     image_size: "2K",
     prompt: buildFramePrompt(edge, ctx.sceneIndex, framePrompt, present, hasOriginal, ctx.shootMode)
   };
-  if (ctx.ar) genParams.aspect_ratio = ctx.ar;
+  if (ctx.genAr) genParams.aspect_ratio = ctx.genAr;
   const genId = `s${ctx.sceneIndex}${tag}_${edge}`;
   nodes.push({
     id: genId,
@@ -12528,7 +12605,7 @@ function buildSeedancePrompt(scene, sceneIndex, present, mode, audio, nativeLine
     if (lines.length > 0)
       parts.push(`Spoken context (do not render as audio): ${lines.map((l) => `"${l}"`).join(" ")}`);
   }
-  const transcript = (scene.transcript_slice ?? []).map((w) => w.text?.trim()).filter(Boolean).join(" ").trim();
+  const transcript = silentClipTranscript(scene, Boolean(nativeLine));
   if (transcript) parts.push(`Transcript: ${loc(transcript)}`);
   const audioLine = seedanceAudioLine(scene, mode, audio, nativeLine);
   if (audioLine) parts.push(audioLine);
@@ -12655,7 +12732,7 @@ function emitSceneClip(i, scene, present, mode, nativeTurn, ambientBroll, frames
     // ambient b-roll beat generates diegetic ambient only; otherwise the clip is silent.
     generate_audio: Boolean(nativeTurn) || ambientBroll
   };
-  if (opts.ar) clipParams.aspect_ratio = opts.ar;
+  if (opts.genAr) clipParams.aspect_ratio = opts.genAr;
   nodes.push({
     id: `s${i}${tag}_clip`,
     type: "video_generate",
@@ -12663,12 +12740,13 @@ function emitSceneClip(i, scene, present, mode, nativeTurn, ambientBroll, frames
     params: clipParams
   });
   const base = `$ref:s${i}${tag}_clip.video`;
-  if (lengths.genDur === lengths.trimTarget) return { ref: base, scene_s: lengths.dur, out };
+  const normDims = opts.outAr && opts.genAr !== opts.outAr ? canvasDims(opts.outAr) : void 0;
+  if (lengths.genDur === lengths.trimTarget && !normDims) return { ref: base, scene_s: lengths.dur, out };
   nodes.push({
     id: `s${i}${tag}_clip_trim`,
     type: "ffmpeg",
     inputs: { clip: base },
-    params: { args: trimArgs(lengths.trimTarget), outputs: { video: { kind: "video", ext: "mp4" } } }
+    params: { args: trimArgs(lengths.trimTarget, 0, normDims), outputs: { video: { kind: "video", ext: "mp4" } } }
   });
   return { ref: `$ref:s${i}${tag}_clip_trim.video`, scene_s: lengths.dur, out };
 }
@@ -12750,7 +12828,7 @@ function slotsForRegion(present, isPresenter) {
   });
 }
 function buildCompositeScene(layout, regions, comp, scene, i, present, mode, nativeTurn, lengths, out, opts, nodes) {
-  const dims = canvasDims(opts.ar);
+  const dims = canvasDims(opts.outAr);
   const presIdx = presenterIndexOf(regions, Boolean(nativeTurn));
   const regionRefs = [];
   let presenterPosition;
@@ -12760,7 +12838,7 @@ function buildCompositeScene(layout, regions, comp, scene, i, present, mode, nat
     const regionSlots = slotsForRegion(present, isPresenter);
     const ctx = {
       sceneIndex: i,
-      ar: opts.ar,
+      genAr: opts.genAr,
       reuse: opts.reuse,
       imageModel: opts.imageModel,
       shootMode: mode,
@@ -12787,7 +12865,13 @@ function buildCompositeScene(layout, regions, comp, scene, i, present, mode, nat
       { first, last },
       lengths,
       null,
-      { ar: opts.ar, videoModel: opts.videoModel, resolution: opts.resolution, nativeLang: opts.nativeLang },
+      {
+        outAr: opts.outAr,
+        genAr: opts.genAr,
+        videoModel: opts.videoModel,
+        resolution: opts.resolution,
+        nativeLang: opts.nativeLang
+      },
       nodes,
       tag
     );
@@ -12864,7 +12948,7 @@ function emitCompositeScene(composite, scene, i, present, mode, nativeTurn, leng
   );
   clips.push(built.clip);
 }
-function emitFlashHold(i, scene, slots, ctx, lengths, out, ar, nodes, clips) {
+function emitFlashHold(i, scene, slots, ctx, lengths, out, outAr, nodes, clips) {
   const frame = buildFrameRef(
     "start",
     scene.start_frame_asset?.url,
@@ -12878,13 +12962,13 @@ function emitFlashHold(i, scene, slots, ctx, lengths, out, ar, nodes, clips) {
     type: "ffmpeg",
     inputs: { frame },
     params: {
-      args: stillHoldArgs(lengths.trimTarget, canvasDims(ar)),
+      args: stillHoldArgs(lengths.trimTarget, canvasDims(outAr)),
       outputs: { video: { kind: "video", ext: "mp4" } }
     }
   });
   clips.push({ ref: `$ref:s${i}_clip.video`, scene_s: lengths.dur, out });
 }
-function emitScreenScene(i, scene, lengths, out, ar, nodes, clips) {
+function emitScreenScene(i, scene, lengths, out, outAr, nodes, clips) {
   const label = commentSafe((scene.summary || scene.start_frame_prompt || "the app screen").slice(0, 120));
   const refId = `s${i}_screen_ref`;
   nodes.push({
@@ -12901,7 +12985,7 @@ function emitScreenScene(i, scene, lengths, out, ar, nodes, clips) {
     type: "ffmpeg",
     inputs: { frame: `$ref:${refId}.asset` },
     params: {
-      args: screenStillArgs(lengths.trimTarget, canvasDims(ar)),
+      args: screenStillArgs(lengths.trimTarget, canvasDims(outAr)),
       outputs: { video: { kind: "video", ext: "mp4" } }
     }
   });
@@ -12943,13 +13027,13 @@ function colorPlateArgs(durationS, dims, color) {
     "{{out.video}}"
   ];
 }
-function emitBrandCardScene(i, lengths, out, ar, color, nodes, clips) {
+function emitBrandCardScene(i, lengths, out, outAr, color, nodes, clips) {
   nodes.push({
     id: `s${i}_clip`,
     type: "ffmpeg",
     inputs: {},
     params: {
-      args: colorPlateArgs(lengths.trimTarget, canvasDims(ar), color),
+      args: colorPlateArgs(lengths.trimTarget, canvasDims(outAr), color),
       outputs: { video: { kind: "video", ext: "mp4" } }
     }
   });
@@ -13081,6 +13165,35 @@ function makePresenterPresent(slots, canonical, opts = {}) {
 var PAUSE_GAP_S = 0.6;
 var SEEDANCE_SAFE_MAX_S = SEEDANCE_DURATIONS.find((d) => d >= 10) ?? 10;
 var PHRASE_MAX_S = SEEDANCE_SAFE_MAX_S;
+var JOIN_DEDUP_MAX_WORDS = 4;
+function joinKey(word) {
+  return word.toLowerCase().replace(/[^\p{L}\p{N}]+/gu, "");
+}
+function joinDialogueTexts(texts) {
+  const out = [];
+  for (const text of texts) {
+    const prev = out[out.length - 1];
+    if (!prev) {
+      out.push(text);
+      continue;
+    }
+    const prevWords = prev.split(/\s+/).filter(Boolean);
+    const nextWords = text.split(/\s+/).filter(Boolean);
+    let overlap = 0;
+    const maxK = Math.min(JOIN_DEDUP_MAX_WORDS, prevWords.length, nextWords.length);
+    for (let k = maxK; k >= 1; k--) {
+      const tail = prevWords.slice(-k).map(joinKey).join(" ");
+      const head = nextWords.slice(0, k).map(joinKey).join(" ");
+      if (tail && tail === head) {
+        overlap = k;
+        break;
+      }
+    }
+    const rest = nextWords.slice(overlap).join(" ");
+    if (rest) out.push(rest);
+  }
+  return out.join(" ");
+}
 function collapseVoiceover(blueprint) {
   const casts = castIdSet(blueprint);
   const cameraOn = onCameraDialogue(blueprint);
@@ -13138,7 +13251,7 @@ function buildPhrases(blueprint, canonical, compositeScenes, presenterPresent, p
       speaker: cur.speaker,
       start_s: cur.start,
       end_s: cur.end,
-      text: cur.texts.join(" "),
+      text: joinDialogueTexts(cur.texts),
       firstScene: cur.firstScene,
       shownScenes,
       presenterShown: shownScenes.length > 0
@@ -13211,7 +13324,7 @@ function emitPhraseClip(phrase, voiceNode, env, nodes, out) {
   const mode = sceneShootMode(anchorScene, present, nativeTurn, env.cameraOn, env.casts);
   const ctx = {
     sceneIndex: anchor,
-    ar: env.ar,
+    genAr: env.genAr,
     reuse: env.reuse,
     imageModel: env.opts.imageModel,
     shootMode: mode,
@@ -13249,7 +13362,7 @@ function emitPhraseClip(phrase, voiceNode, env, nodes, out) {
     ...videoResolutionParam(env.opts.videoModel, env.opts.resolution),
     generate_audio: true
   };
-  if (env.ar) clipParams.aspect_ratio = env.ar;
+  clipParams.aspect_ratio = env.genAr;
   nodes.push({
     id: `s${anchor}_clip`,
     type: "video_generate",
@@ -13363,7 +13476,8 @@ function emitCompositeInTimeline(composite, scene, i, isLast, env, canonical, en
     lengths,
     lengths.out,
     {
-      ar: env.ar,
+      outAr: env.outAr,
+      genAr: env.genAr,
       reuse: env.reuse,
       imageModel: env.opts.imageModel,
       videoModel: env.opts.videoModel,
@@ -13433,23 +13547,23 @@ function emitBrollScene(scene, i, isLast, env, nodes, out, prevEndFrame) {
   const lengths = sceneTiming(scene, isLast, void 0);
   const ctx = {
     sceneIndex: i,
-    ar: env.ar,
+    genAr: env.genAr,
     reuse: env.reuse,
     imageModel: env.opts.imageModel,
     shootMode: mode,
     ingestCache: env.ingestCache
   };
   if (!env.reuse && sceneIsFullScreenUi(scene, present)) {
-    emitScreenScene(i, scene, lengths, lengths.out, env.ar, nodes, out.clips);
+    emitScreenScene(i, scene, lengths, lengths.out, env.outAr, nodes, out.clips);
     return void 0;
   }
   const isCta = scene.narrative_role?.trim() === "cta" || isLast;
   if (!env.reuse && sceneIsBrandCard(scene, present, isCta)) {
-    emitBrandCardScene(i, lengths, lengths.out, env.ar, brandPlateColor(env.blueprint), nodes, out.clips);
+    emitBrandCardScene(i, lengths, lengths.out, env.outAr, brandPlateColor(env.blueprint), nodes, out.clips);
     return void 0;
   }
   if (!ambientBroll && lengths.dur <= FLASH_HOLD_MAX_S) {
-    emitFlashHold(i, scene, env.slots, ctx, lengths, lengths.out, env.ar, nodes, out.clips);
+    emitFlashHold(i, scene, env.slots, ctx, lengths, lengths.out, env.outAr, nodes, out.clips);
     return void 0;
   }
   const first = scene.continues_previous && prevEndFrame ? prevEndFrame : buildFrameRef(
@@ -13478,7 +13592,13 @@ function emitBrollScene(scene, i, isLast, env, nodes, out, prevEndFrame) {
     { first, last },
     { dur: lengths.dur, trimTarget: lengths.trimTarget, genDur: lengths.genDur },
     lengths.out,
-    { ar: env.ar, videoModel: env.opts.videoModel, resolution: env.opts.resolution, nativeLang: env.ttsLanguageCode },
+    {
+      outAr: env.outAr,
+      genAr: env.genAr,
+      videoModel: env.opts.videoModel,
+      resolution: env.opts.resolution,
+      nativeLang: env.ttsLanguageCode
+    },
     nodes
   );
   if (ambientBroll) {
@@ -13506,11 +13626,13 @@ function buildTimeline(blueprint, slots, opts, nodes) {
   }
   const canonical = collapseVoiceover(blueprint);
   const ensureVoiceNode = makeVoiceFactory(blueprint, canonical, nodes);
+  const aspect = resolveAspect(blueprint.source?.aspect_ratio, opts.aspect, genAspectsFor(opts.videoModel));
   const env = {
     blueprint,
     slots,
     opts,
-    ar: aspectRatioParam(blueprint),
+    outAr: aspect.outAr,
+    genAr: aspect.genAr,
     reuse,
     cameraOn: onCameraDialogue(blueprint),
     casts: castIdSet(blueprint),
@@ -13562,7 +13684,8 @@ function buildTimeline(blueprint, slots, opts, nodes) {
     }
     const slice = out.sceneSlice.get(i);
     if (slice) {
-      const whole = slice.offset === 0 && Math.abs(slice.len - slice.clipDur) <= 0.05;
+      const normDims = env.genAr !== env.outAr ? canvasDims(env.outAr) : void 0;
+      const whole = slice.offset === 0 && Math.abs(slice.len - slice.clipDur) <= 0.05 && !normDims;
       if (whole) {
         out.clips.push({ ref: slice.clipRef, scene_s: slice.len, out: null });
       } else {
@@ -13570,7 +13693,10 @@ function buildTimeline(blueprint, slots, opts, nodes) {
           id: `s${i}_seg`,
           type: "ffmpeg",
           inputs: { clip: slice.clipRef },
-          params: { args: trimArgs(slice.len, slice.offset), outputs: { video: { kind: "video", ext: "mp4" } } }
+          params: {
+            args: trimArgs(slice.len, slice.offset, normDims),
+            outputs: { video: { kind: "video", ext: "mp4" } }
+          }
         });
         out.clips.push({ ref: `$ref:s${i}_seg.video`, scene_s: slice.len, out: null });
       }
@@ -13687,6 +13813,8 @@ function collectCaptions(blueprint) {
     }) : [];
   }).sort((a, b) => a.at - b.at);
 }
+var OVERLAY_REDETECT_GAP_S = 2;
+var OVERLAY_MIN_DUR_S = 0.5;
 function mergeCaptions(blueprint) {
   const byText = /* @__PURE__ */ new Map();
   for (const e of collectCaptions(blueprint)) {
@@ -13698,7 +13826,7 @@ function mergeCaptions(blueprint) {
   for (const arr of byText.values()) {
     let cur = null;
     for (const e of arr) {
-      if (cur && e.at <= cur.end + 0.35) cur.end = Math.max(cur.end, e.end);
+      if (cur && e.at <= cur.end + OVERLAY_REDETECT_GAP_S) cur.end = Math.max(cur.end, e.end);
       else {
         cur = { ...e };
         merged.push(cur);
@@ -13787,10 +13915,28 @@ function buildOverlayHtml(input) {
   ];
   const ovParts = mergeCaptions(blueprint).map((e) => overlayElement(e.ov, e.at, Math.round((e.end - e.at) * 1e3) / 1e3)).filter(Boolean);
   if (ovParts.length > 0) blocks.push(ovParts.join("\n"));
+  const seenFloats = /* @__PURE__ */ new Map();
+  const keepFloat = (fe, sceneStart) => {
+    const at = fe.appears_at_s ?? sceneStart;
+    const dur = fe.duration_s ?? 2.5;
+    if (dur < OVERLAY_MIN_DUR_S) return false;
+    const key = [
+      (fe.kind ?? "element").toLowerCase(),
+      (fe.brand_name || fe.what_it_represents || fe.description || "").toLowerCase().trim(),
+      positionClass(fe.position)
+    ].join("|");
+    const lastEnd = seenFloats.get(key);
+    if (lastEnd !== void 0 && at <= lastEnd + OVERLAY_REDETECT_GAP_S) {
+      seenFloats.set(key, Math.max(lastEnd, at + dur));
+      return false;
+    }
+    seenFloats.set(key, at + dur);
+    return true;
+  };
   for (const scene of blueprint.scenes) {
     const sceneStart = scene.start_s ?? 0;
     const floats = z9.array(FloatingElement).safeParse(scene.floating_elements ?? []);
-    const parts = (floats.success ? floats.data.map((fe) => floatingStub(fe, sceneStart)) : []).filter(Boolean);
+    const parts = (floats.success ? floats.data.filter((fe) => keepFloat(fe, sceneStart)).map((fe) => floatingStub(fe, sceneStart)) : []).filter(Boolean);
     const pip = uiPipStub(scene);
     if (pip) parts.push(pip);
     if (parts.length > 0) blocks.push(parts.join("\n"));
@@ -13994,12 +14140,16 @@ function scaffoldVideoCanvas(input, elementsInput, opts) {
     });
     videoNode = "final";
   }
+  const todo = {
+    ...buildVideoTodo(videoReport(input, elementsInput), overlays.length, floating.length, opts, blueprint),
+    ...aspectRemapTodo(resolveAspect(blueprint.source?.aspect_ratio, opts.aspect, genAspectsFor(opts.videoModel)))
+  };
   return {
     schema: "baker-canvas/1",
     metadata: {
       name: "video reproduction",
       description: VIDEO_GUIDE,
-      todo: buildVideoTodo(videoReport(input, elementsInput), overlays.length, floating.length, opts, blueprint),
+      todo,
       // The timing plan `baker canvas validate` checks before any billed render:
       // sequenced voiceover turns (no overlap), audio ≈ video length, and which
       // scenes must be lip-synced.
@@ -14360,6 +14510,27 @@ async function stageCaptions(outDir, transcript) {
   await cp(SHIPPED_CAPTIONS_DIR, compositionPath, { recursive: true });
   return { compositionPath };
 }
+function patchCompositionMeta(metaJson, dims) {
+  try {
+    const meta = JSON.parse(metaJson);
+    if (!meta || typeof meta !== "object" || Array.isArray(meta)) return metaJson;
+    return `${JSON.stringify({ ...meta, width: dims.w, height: dims.h }, null, 2)}
+`;
+  } catch {
+    return metaJson;
+  }
+}
+function patchCompositionHtml(html, dims) {
+  return html.replace(/(<meta\s+name="viewport"\s+content="width=)\d+(,\s*height=)\d+(")/i, `$1${dims.w}$2${dims.h}$3`).replace(/(width:\s*)\d+(px;\s*height:\s*)\d+(px;)/i, `$1${dims.w}$2${dims.h}$3`).replace(/(data-width=")\d+(")/i, `$1${dims.w}$2`).replace(/(data-height=")\d+(")/i, `$1${dims.h}$2`);
+}
+async function stampCompositionDims(compositionDir, dims) {
+  const metaPath = path9.join(compositionDir, "meta.json");
+  const rawMeta = await readFile6(metaPath, "utf8");
+  await writeFile2(metaPath, patchCompositionMeta(rawMeta, dims), "utf8");
+  const htmlPath = path9.join(compositionDir, "index.html");
+  const rawHtml = await readFile6(htmlPath, "utf8");
+  await writeFile2(htmlPath, patchCompositionHtml(rawHtml, dims), "utf8");
+}
 function parseElements2(raw) {
   const parsed = JSON.parse(raw);
   if (Array.isArray(parsed)) return parsed;
@@ -14512,6 +14683,10 @@ var scaffoldVideoCommand = defineCommand85({
     "select-model": { type: "string", description: "Override the text_generate model id for element selection" },
     "image-model": { type: "string", description: "Override the image_generate model id for frames" },
     "video-model": { type: "string", description: "Override the video_generate model id for clips" },
+    aspect: {
+      type: "string",
+      description: `Output aspect ratio (e.g. "4:5"). Default: the source video's detected ratio. When the video model can't generate the ratio (Seedance has no 4:5) clips generate at the nearest supported ratio and are center-cropped onto the output canvas.`
+    },
     resolution: {
       type: "string",
       description: `Output resolution for generated clips (e.g. "1080p"). Default 1080p \u2014 the highest the video model supports \u2014 so clips keep the keyframe sharpness instead of the model's low default.`
@@ -14545,8 +14720,26 @@ var scaffoldVideoCommand = defineCommand85({
     const annotated = annotateBlueprintWithElements(blueprint, elements);
     await writeFile2(blueprintPath, `${JSON.stringify(annotated, null, 2)}
 `, "utf8");
+    let aspect;
+    try {
+      aspect = resolveAspect(
+        blueprint.source?.aspect_ratio,
+        args.aspect ? String(args.aspect) : void 0,
+        genAspectsFor(videoModel)
+      );
+    } catch (e) {
+      return fail2("aspect", e instanceof Error ? e.message : String(e));
+    }
+    const outDims = canvasDims(aspect.outAr);
+    if (aspect.remapped) {
+      process.stderr.write(
+        `\u26A0\uFE0F  ${aspect.outAr} output \u2192 clips generate at ${aspect.genAr} (the video model's nearest ratio) and are center-cropped to ${aspect.outAr} (${outDims.w}\xD7${outDims.h}) \u2014 keep subjects centered in frame prompts.
+`
+      );
+    }
     const compositionDest = path9.join(outDir, "video-overlay-composition");
     await cp(SHIPPED_COMPOSITION_DIR, compositionDest, { recursive: true });
+    await stampCompositionDims(compositionDest, outDims);
     const indexPath = path9.join(compositionDest, "index.html");
     const overlayHtml = buildOverlayHtml(blueprint);
     const indexHtml = await readFile6(indexPath, "utf8");
@@ -14559,6 +14752,7 @@ var scaffoldVideoCommand = defineCommand85({
     }
     await writeFile2(indexPath, injected, "utf8");
     const captions = await stageCaptions(outDir, transcript);
+    if (captions.compositionPath) await stampCompositionDims(captions.compositionPath, outDims);
     const opts = {
       imageModel,
       videoModel,
@@ -14567,6 +14761,7 @@ var scaffoldVideoCommand = defineCommand85({
       blueprintPath: path9.relative(outDir, blueprintPath),
       frames,
       ambient: Boolean(args.ambient),
+      ...args.aspect ? { aspect: String(args.aspect) } : {},
       ...args.resolution ? { resolution: String(args.resolution) } : {}
     };
     let canvas;