@koda-sl/baker-cli 0.301.0-dev.50ab1c583 → 0.305.0-dev.2c645fe6b

package/dist/cli.js

@@ -22,6 +22,8 @@ import {
   YtDlpError,
   bestSupportedQuality,
   canonicalUrl,
+  clipAudioMode,
+  clipDurationsFor,
   clipParamRecipe,
   clipProfileFor,
   collectAssetRefLikes,
@@ -63,7 +65,7 @@ import {
   ulid,
   validateCanvasDeep,
   ytDlpBlockSignal
-} from "./chunk-SSXER76B.js";
+} from "./chunk-7TL4VPK5.js";
 import {
   csvOrJson,
   daysAgoIso,
@@ -2380,7 +2382,6 @@ var IMAGE_ASSET_LINK_FIELD_TYPES = /* @__PURE__ */ new Set([
   "PORTRAIT_MARKETING_IMAGE",
   "AD_IMAGE"
 ]);
-var IMAGE_ASSET_LINK_MESSAGE = "Google refuses the LINK, not the upload \u2014 so do not stop at this refusal, finish the handover. `assets create` with an image DOES publish: the image lands in the advertiser's Google Ads asset library and is theirs to use. Only this link is refused (MARKETING_IMAGE is a Performance Max asset-group slot \u2014 use `asset-groups attach`; AD_IMAGE comes back UNSUPPORTED_FIELD_TYPE at every level, on every API version). So: stage the image assets anyway, drop this link, and tell the user the images are already in their Google Ads account and only need adding to the campaign under Assets \u2014 name them. Baker stages the Task for that step itself, alongside the image, so it is already in this Session's changes \u2014 do not stage a second one with `baker actions`. Never tell them to upload the images; they are uploaded. The one image a Search campaign does take through a link is the advertiser's logo \u2014 BUSINESS_LOGO at campaign or customer level, on a verified account.";
 function isImageAssetLinkFieldType(fieldType) {
   return IMAGE_ASSET_LINK_FIELD_TYPES.has(fieldType);
 }
@@ -3384,7 +3385,6 @@ var assetLinkAttachSchema = z5.object({
   if (value.fieldType === "BUSINESS_LOGO" && value.level === "adGroup") {
     ctx.addIssue({ code: z5.ZodIssueCode.custom, path: ["fieldType"], message: BUSINESS_LOGO_LEVEL_MESSAGE });
   } else if (isImageAssetLinkFieldType(value.fieldType)) {
-    ctx.addIssue({ code: z5.ZodIssueCode.custom, path: ["fieldType"], message: IMAGE_ASSET_LINK_MESSAGE });
   } else if (!ASSET_FIELD_TYPES.includes(value.fieldType)) {
     ctx.addIssue({
       code: z5.ZodIssueCode.custom,
@@ -4464,7 +4464,13 @@ var avatarsCreateRequestSchema = z10.object({
   sourceImageIds: z10.array(z10.string().min(1)).max(6).optional(),
   /** The company's statement that it may use this person's likeness in ads. */
   likenessConfirmed: z10.boolean().optional(),
-  chatId: z10.string().optional()
+  chatId: z10.string().optional(),
+  /**
+   * Who picks this avatar's face. `ask` (the default) means a picker is about to open,
+   * so the sheet waits for the chosen portrait instead of being fused from one nobody
+   * saw — two renders that were always going to be thrown away.
+   */
+  faceChoice: z10.enum(["ask", "auto"]).optional()
 }).refine((body) => !body.sourceImageIds?.length || body.likenessConfirmed === true, {
   message: "likenessConfirmed is required when building an avatar from photographs of a real person",
   path: ["likenessConfirmed"]
@@ -16592,8 +16598,8 @@ function imageAssetHints(payload) {
     return [];
   }
   return [
-    "This image publishes into the advertiser's Google Ads asset library \u2014 that half works. What Google refuses is the link that puts it on a Search campaign, so do not stage an `assets attach` for it.",
-    "Finish the handover rather than reporting a limitation: tell the user the images are already in their Google Ads account and only need adding to the campaign under Assets, and name the ones you staged. Baker stages the Task for that step itself the moment you stage the image, so it is already in this chat's changes \u2014 do not stage a second one with `baker actions`. Never tell them to upload images they already have. (A Performance Max asset group is different \u2014 images do link there, with `asset-groups attach`.)"
+    "This image publishes into the advertiser's Google Ads asset library \u2014 that half works. What Google refuses is the link that puts it on a Search campaign, so the `assets attach` for it never stores: it is skipped with the reason, and staging it anyway is what tells Baker which campaign these images belong on.",
+    "Finish the handover rather than reporting a limitation: tell the user the images are already in their Google Ads account and only need adding to the campaign under Assets, and name the ones you staged. Baker stages the Task for that step itself the moment you stage the image, so it is already in this chat's changes \u2014 do not stage a second one with `baker actions`. Stage the `assets attach` too if you know which campaign they belong on: it is skipped rather than stored, and it is what puts the campaign's name in that Task. Never tell them to upload images they already have. (A Performance Max asset group is different \u2014 images do link there, with `asset-groups attach`.)"
   ];
 }
 var assetsUpdateCommand = defineCommand31({
@@ -16636,7 +16642,7 @@ var assetsUpdateCommand = defineCommand31({
 var assetsCommand = defineCommand31({
   meta: {
     name: "assets",
-    description: "Stage asset create/update + asset-link attach/detach (via --file). For a high-performance Search campaign add \u22654 sitelinks, \u22653 callouts and \u22651 structured snippet per campaign. Image extensions are the one thing you cannot finish here \u2014 `create` uploads the image into the advertiser's account, but Google refuses the link that puts it on a Search campaign, so hand that one step over rather than telling them to upload anything. The image a Search campaign does take through a link is the advertiser's logo \u2014 BUSINESS_LOGO, at campaign or customer level, on a verified account."
+    description: "Stage asset create/update + asset-link attach/detach (via --file). For a high-performance Search campaign add \u22654 sitelinks, \u22653 callouts and \u22651 structured snippet per campaign. Image extensions are the one thing you cannot finish here \u2014 `create` uploads the image into the advertiser's account, but Google refuses the link that puts it on a Search campaign, so hand that one step over rather than telling them to upload anything (stage the `attach` regardless: it is skipped, and it is what records the handover Task). The image a Search campaign does take through a link is the advertiser's logo \u2014 BUSINESS_LOGO, at campaign or customer level, on a verified account."
   },
   subCommands: {
     create: fileCreateCommand(
@@ -16647,7 +16653,7 @@ var assetsCommand = defineCommand31({
     ),
     update: assetsUpdateCommand,
     attach: fileCreateCommand("attach", "google.assetLink.attach", "Attach an asset to a campaign/adGroup/customer", [
-      "Image extensions are not attachable here \u2014 Google refuses AD_IMAGE and every marketing-image field type at every level, on every API version. The upload is not refused, though: stage the images with `assets create`, then tell the user they are already in their Google Ads account and only need adding to the campaign under Assets. Baker stages that Task by itself when the image is staged. Two images that DO work: the advertiser's logo through this command (`BUSINESS_LOGO`, campaign or customer level, on a verified account), and Performance Max asset-group images, which are a different mechanism (`asset-groups attach`)."
+      "Image extensions cannot be linked \u2014 Google refuses AD_IMAGE and every marketing-image field type at every level, on every API version \u2014 but stage the link anyway: it is skipped rather than stored, the rest of a batch is unaffected, and it is what records the handover Task naming those images and that campaign. That is the ONLY record when the images are already in the account and you are uploading nothing. Then tell the user they are already in their Google Ads account and only need adding to the campaign under Assets; do not stage a second Task with `baker actions`. Two images that DO link: the advertiser's logo through this command (`BUSINESS_LOGO`, campaign or customer level, on a verified account), and Performance Max asset-group images, which are a different mechanism (`asset-groups attach`)."
     ]),
     detach: statusCommand2("google.assetLink.detach", "asset link")
   }
@@ -30245,7 +30251,10 @@ var createCommand2 = defineCommand96({
         ...sourceImageIds ? { sourceImageIds, likenessConfirmed: true } : {},
         ...args["voice-id"] ? { voiceId: args["voice-id"] } : {},
         ...args["voice-description"] ? { voiceDescription: args["voice-description"] } : {},
-        ...chatIdFromEnv() ? { chatId: chatIdFromEnv() } : {}
+        ...chatIdFromEnv() ? { chatId: chatIdFromEnv() } : {},
+        // Sent so the backend can hold the sheet until the chosen portrait exists —
+        // otherwise it fuses one from a face nobody picked and pays for it twice.
+        faceChoice
       };
       const data = await writeAvatars("/api/avatars", body);
       writeJson({ ok: true, data, hints: creationHints({ ...data, faceChoice }, body.profile) });
@@ -33462,7 +33471,7 @@ function sameSpeakerTrackEntries(segments) {
     };
   });
 }
-function buildPerSpeakerVoiceConversion(segments, totalMs, nodes) {
+function buildPerSpeakerVoiceConversion(segments, totalMs, nodes, keepRenderedVoice = false) {
   const bySpeaker = /* @__PURE__ */ new Map();
   for (const seg of segments) {
     const arr = bySpeaker.get(seg.voiceNode) ?? [];
@@ -33471,7 +33480,7 @@ function buildPerSpeakerVoiceConversion(segments, totalMs, nodes) {
   }
   const tracks = [];
   for (const [voiceNode, segs] of bySpeaker) {
-    const trackId = `${voiceNode}_track`;
+    const trackId = keepRenderedVoice ? `${voiceNode}_conv` : `${voiceNode}_track`;
     const convId = `${voiceNode}_conv`;
     const sorted = [...segs].sort((a, b) => a.start_s - b.start_s);
     const mixInputs = {};
@@ -33487,12 +33496,14 @@ function buildPerSpeakerVoiceConversion(segments, totalMs, nodes) {
         total_ms: totalMs
       }
     });
-    nodes.push({
-      id: convId,
-      type: "audio_voice_convert",
-      inputs: { audio: `$ref:${trackId}.audio`, voice_ref: `$ref:${voiceNode}.voice_id` },
-      params: { model: FIXED_VOICE_CONVERT_MODEL, voice: "{{voice_ref}}" }
-    });
+    if (!keepRenderedVoice) {
+      nodes.push({
+        id: convId,
+        type: "audio_voice_convert",
+        inputs: { audio: `$ref:${trackId}.audio`, voice_ref: `$ref:${voiceNode}.voice_id` },
+        params: { model: FIXED_VOICE_CONVERT_MODEL, voice: "{{voice_ref}}" }
+      });
+    }
     tracks.push({ slot: convId, ref: `$ref:${convId}.audio`, start_s: 0, duration_s: totalMs / 1e3, kind: "vo" });
   }
   return tracks;
@@ -34632,7 +34643,8 @@ function buildTimeline(blueprint, slots, opts, nodes) {
     graphicScenes: fullFrameGraphicScenes(blueprint),
     surfaceIngests: /* @__PURE__ */ new Map(),
     ttsLanguageCode: deriveTtsLanguageCode(blueprint),
-    spokenAccent: blueprint.global?.voiceover?.accent?.trim() || void 0
+    spokenAccent: blueprint.global?.voiceover?.accent?.trim() || void 0,
+    keepRenderedVoice: opts.keepRenderedVoice
   };
   const out = {
     routes: /* @__PURE__ */ new Map(),
@@ -34689,7 +34701,7 @@ function buildTimeline(blueprint, slots, opts, nodes) {
     prevEndFrame = emitBrollScene(scene, i, i === lastIndex, env, nodes, out, prevEndFrame);
   });
   const totalMs = Math.round(env.clock.totalS * 1e3);
-  out.voTracks.push(...buildPerSpeakerVoiceConversion(out.nativeSegments, totalMs, nodes));
+  out.voTracks.push(...buildPerSpeakerVoiceConversion(out.nativeSegments, totalMs, nodes, env.keepRenderedVoice));
   return {
     clips: out.clips,
     voTracks: out.voTracks,
@@ -35127,6 +35139,18 @@ function emitCaptionChain(blueprint, tracks, clock, opts, videoRef, nodes) {
   });
   return { videoRef: "$ref:captions.video", videoNode: "captions" };
 }
+function dropUncastVoices(nodes) {
+  const referenced = /* @__PURE__ */ new Set();
+  for (const node of nodes) {
+    for (const ref of Object.values(node.inputs ?? {})) {
+      if (typeof ref === "string" && ref.startsWith("$ref:")) referenced.add(ref.slice(5).split(".")[0] ?? "");
+    }
+  }
+  for (let i = nodes.length - 1; i >= 0; i--) {
+    const node = nodes[i];
+    if (node && node.type === "voice_select" && !referenced.has(node.id)) nodes.splice(i, 1);
+  }
+}
 function scaffoldVideoCanvas(input, elementsInput, opts) {
   const blueprint = VideoBlueprint.parse(input);
   injectHookPhysicality(blueprint);
@@ -35240,6 +35264,7 @@ function scaffoldVideoCanvas(input, elementsInput, opts) {
     ...aspectRemapTodo(resolveAspect(blueprint.source?.aspect_ratio, opts.aspect, genAspectsFor(opts.videoModel))),
     model_constraints: buildModelConstraints()
   };
+  dropUncastVoices(nodes);
   return {
     schema: "baker-canvas/1",
     metadata: {
@@ -36361,6 +36386,14 @@ var RENDER_REVIEW_MODEL = "gemini-pro-latest";
 var RENDER_REVIEW_FRAMES = 12;
 var RENDER_REVIEW_WIDTH = 1024;
 var RENDER_WATCH_MODEL = "gemini-3.7-flash";
+function usageFrom(model, body) {
+  const meta = body?.usageMetadata;
+  if (!meta) return null;
+  const inputTokens = meta.promptTokenCount ?? 0;
+  const outputTokens = meta.candidatesTokenCount ?? 0;
+  if (inputTokens === 0 && outputTokens === 0) return null;
+  return { model, inputTokens, outputTokens };
+}
 var WATCH_MAX_BYTES = 60 * 1024 * 1024;
 function sceneAt(windows, atS) {
   return windows.find((w) => atS >= w.startS && atS < w.endS)?.file;
@@ -36622,6 +36655,12 @@ function videoPathFromOutput(output) {
   }
   return found;
 }
+function parseVolume(volumedetect, key) {
+  const raw = new RegExp(`${key}:\\s*(-?(?:inf|[\\d.]+)) dB`).exec(volumedetect)?.[1];
+  if (raw === void 0) return Number.NaN;
+  if (raw === "-inf") return Number.NEGATIVE_INFINITY;
+  return Number.parseFloat(raw);
+}
 function classifyAudio(volumedetect, hasAudioStream) {
   if (!hasAudioStream) {
     return [
@@ -36631,9 +36670,17 @@ function classifyAudio(volumedetect, hasAudioStream) {
       }
     ];
   }
-  const mean = Number.parseFloat(volumedetect.match(/mean_volume:\s*(-?[\d.]+) dB/)?.[1] ?? "");
-  const max = Number.parseFloat(volumedetect.match(/max_volume:\s*(-?[\d.]+) dB/)?.[1] ?? "");
+  const mean = parseVolume(volumedetect, "mean_volume");
+  const max = parseVolume(volumedetect, "max_volume");
   const findings = [];
+  if (mean === Number.NEGATIVE_INFINITY || max === Number.NEGATIVE_INFINITY) {
+    return [
+      {
+        severity: "blocking",
+        what: "The finished video has an audio track, but it is completely silent \u2014 digital zero from start to end. An ad that plays silent in a feed is not an ad."
+      }
+    ];
+  }
   if (Number.isFinite(mean) && mean < -30) {
     findings.push({
       severity: "warning",
@@ -36691,7 +36738,7 @@ async function videoDuration(video) {
     return null;
   }
 }
-async function askVideo(video, prompt) {
+async function askVideo(video, prompt, spent) {
   const key = frameVisionKey();
   if (!key) return null;
   try {
@@ -36714,13 +36761,15 @@ async function askVideo(video, prompt) {
     );
     if (!response.ok) return null;
     const body = await response.json();
+    const usage = usageFrom(RENDER_WATCH_MODEL, body);
+    if (usage) spent(usage);
     const text2 = body.candidates?.[0]?.content?.parts?.map((p) => p.text ?? "").join("");
     return text2 ? parseReviewJson(text2) : null;
   } catch {
     return null;
   }
 }
-async function ask(images, prompt) {
+async function ask(images, prompt, spent) {
   const key = frameVisionKey();
   if (!key) return null;
   const parts = [{ text: prompt }];
@@ -36737,6 +36786,8 @@ async function ask(images, prompt) {
     );
     if (!response.ok) return null;
     const body = await response.json();
+    const usage = usageFrom(RENDER_REVIEW_MODEL, body);
+    if (usage) spent(usage);
     const text2 = body.candidates?.[0]?.content?.parts?.map((p) => p.text ?? "").join("");
     return text2 ? parseReviewJson(text2) : null;
   } catch {
@@ -36747,6 +36798,7 @@ async function reviewRenderedVideo(opts) {
   if (!hasFrameVisionKey()) return null;
   const duration = await videoDuration(opts.video);
   if (duration === null) return null;
+  const spent = (usage) => opts.onSpend?.(usage);
   const dir = await mkdtemp(path15.join(tmpdir(), "baker-review-"));
   try {
     const grabbed = [];
@@ -36761,17 +36813,22 @@ async function reviewRenderedVideo(opts) {
     if (grabbed.length === 0) return null;
     const reel = await ask(
       grabbed.map((f) => f.image),
-      buildReelReviewPrompt({ subject: opts.subject, script: opts.script })
+      buildReelReviewPrompt({ subject: opts.subject, script: opts.script }),
+      spent
     );
     const framePrompt = buildFrameReviewPrompt({ subject: opts.subject });
     const perFrame = await mapWithLimit(grabbed, REVIEW_CONCURRENCY, async (f) => ({
       at: f.at,
-      json: await ask([f.image], framePrompt)
+      json: await ask([f.image], framePrompt, spent)
     }));
     const unread = perFrame.filter((f) => f.json === null).length;
     if (unread === perFrame.length && reel === null) return null;
-    const watch = await askVideo(opts.video, buildWatchReviewPrompt({ market: opts.market, script: opts.script }));
-    const fidelity = opts.brief?.trim() ? await askVideo(opts.video, buildFidelityPrompt(opts.brief)) : null;
+    const watch = await askVideo(
+      opts.video,
+      buildWatchReviewPrompt({ market: opts.market, script: opts.script }),
+      spent
+    );
+    const fidelity = opts.brief?.trim() ? await askVideo(opts.video, buildFidelityPrompt(opts.brief), spent) : null;
     const findings = [
       ...classifyFrameFindings(perFrame, opts.windows ?? []),
       ...classifyReelFindings(reel),
@@ -36781,12 +36838,15 @@ async function reviewRenderedVideo(opts) {
       // Measured, not judged: a model looking at frames cannot hear the mix.
       ...await measureAudio(opts.video)
     ];
-    if (unread > 0) {
-      findings.push({
-        severity: "warning",
-        what: `${unread} of ${perFrame.length} frames could not be reviewed \u2014 that part of the video is unchecked.`
-      });
-    }
+    findings.push(
+      ...unreviewedNotes({
+        watched: watch !== null,
+        fidelityAsked: Boolean(opts.brief?.trim()),
+        fidelityRead: fidelity !== null,
+        unread,
+        frames: perFrame.length
+      })
+    );
     return findings;
   } finally {
     await rm5(dir, { recursive: true, force: true });
@@ -36810,6 +36870,28 @@ async function sceneWindowsBeside(canvasPath) {
   }
   return windows;
 }
+function unreviewedNotes(state) {
+  const notes = [];
+  if (!state.watched) {
+    notes.push({
+      severity: "warning",
+      what: "Nobody watched the finished video end to end \u2014 the accent, the voice against the mouth and the cuts are unchecked. The frame-by-frame review still ran."
+    });
+  }
+  if (state.fidelityAsked && !state.fidelityRead) {
+    notes.push({
+      severity: "warning",
+      what: "The video was not checked against what you asked for \u2014 only against whether it is well made. A well-made ad of the wrong brief passes every other check here."
+    });
+  }
+  if (state.unread > 0) {
+    notes.push({
+      severity: "warning",
+      what: `${state.unread} of ${state.frames} frames could not be reviewed \u2014 that part of the video is unchecked.`
+    });
+  }
+  return notes;
+}
 
 // src/commands/canvas/run.ts
 var runCommand = defineCommand107({
@@ -37124,10 +37206,12 @@ ${describeRewrites(healed.rewrites)}
 `));
     }
     const reviewable = videoPathFromOutput(result.output);
+    const reviewUsage = [];
     const review = reviewable ? await reviewRenderedVideo({
       video: reviewable,
       windows: await sceneWindowsBeside(filePath),
-      brief: opts.brief?.trim() || void 0
+      brief: opts.brief?.trim() || void 0,
+      onSpend: (usage) => reviewUsage.push(usage)
     }) : null;
     const hints2 = reviewable ? review === null ? [RENDER_REVIEW_UNAVAILABLE] : review.length === 0 ? ["Reviewed the finished video frame by frame \u2014 nothing to report."] : [
       `Reviewed the finished video and found ${review.length} thing${review.length === 1 ? "" : "s"} to fix. Show the user what is wrong before you show them the ad.`,
@@ -37141,7 +37225,12 @@ ${describeRewrites(healed.rewrites)}
         // `review` is null only when nobody looked (no key, no ffmpeg, unreadable video);
         // an empty array is a render that was reviewed and came back clean. The record
         // has to keep those apart.
-        buildRunRecord(result, { ...recordMeta, ...review ? { review } : {} }, progress?.planInfo(), finalLabels)
+        buildRunRecord(
+          result,
+          { ...recordMeta, ...review ? { review } : {}, ...reviewUsage.length > 0 ? { reviewUsage } : {} },
+          progress?.planInfo(),
+          finalLabels
+        )
       );
     }
     process.stdout.write(
@@ -38252,6 +38341,12 @@ function pinnedModelHint(input) {
   if (!speaks) return null;
   return `You pinned --model ${input.model} on a speaking avatar. Without it this clip routes to ${DEFAULT_VIDEO_GENERATE_MODEL}, which was chosen for exactly this case: the better picture, both orientations, and audio always on. Drop --model unless the user asked for a specific one \u2014 and note ${input.model} may not take the length or the reference you are counting on.`;
 }
+function silentModelHint(model, prompt) {
+  if (!model || !prompt) return null;
+  if (clipAudioMode(model) !== "never") return null;
+  if (spokenWordsIn(prompt) === 0) return null;
+  return `${model} renders no audio \u2014 the line in this prompt will be mouthed and never heard. For a shot where someone speaks use ${DEFAULT_VIDEO_GENERATE_MODEL} (audio always on), or alibaba/wan-3.0 past 10 seconds. Keep this model for shots with no voice in them.`;
+}
 
 // src/commands/canvas/scaffold-ad.ts
 import { copyFile, cp, mkdir as mkdir7, readFile as readFile16, stat as stat4, writeFile as writeFile9 } from "fs/promises";
@@ -38461,6 +38556,7 @@ var CURRENCY_BY_MARKET = {
   "the united kingdom": "\xA3 (pounds sterling)",
   "the united states": "$ (US dollars)"
 };
+var MARKERS_TO_NAME_A_PLACE = 3;
 var MARKET_BY_LANGUAGE = {
   es: "Spain",
   "es-es": "Spain",
@@ -38499,6 +38595,9 @@ var SCRIPT_LANGUAGE_MARKERS = [
   ["english", /\b(the|a|an|of|to|for|with|your|more|already|but|because|until|without)\b/gi]
 ];
 function detectScriptLanguage(lines) {
+  return bestScriptLanguage(lines)?.name;
+}
+function bestScriptLanguage(lines) {
   const text2 = lines.join(" ");
   if (text2.trim().length < 12) return void 0;
   let best = null;
@@ -38506,7 +38605,7 @@ function detectScriptLanguage(lines) {
     const hits = (text2.match(re2) ?? []).length;
     if (hits > 0 && (!best || hits > best.hits)) best = { name, hits };
   }
-  return best?.name;
+  return best ?? void 0;
 }
 function voiceLanguageFor(spec) {
   const declared = spec.voice?.language?.trim().toLowerCase();
@@ -38519,12 +38618,19 @@ function voiceLanguageFor(spec) {
 function marketFor(spec) {
   if (spec.market) return spec.market;
   const lang = spec.voice?.language?.trim().toLowerCase();
-  return lang ? MARKET_BY_LANGUAGE[lang] : void 0;
-}
-function beatSeconds(line) {
+  const declared = lang ? MARKET_BY_LANGUAGE[lang] : void 0;
+  if (declared) return declared;
+  const read = bestScriptLanguage(spec.beats.map((b) => b.say));
+  if (!read || read.hits < MARKERS_TO_NAME_A_PLACE) return void 0;
+  const detected = read.name;
+  if (detected.toLowerCase() === "english") return void 0;
+  const code = detected ? Object.entries(VOICE_LANGUAGE_NAME).find(([, name]) => name.toLowerCase() === detected.toLowerCase())?.[0] : void 0;
+  return code ? MARKET_BY_LANGUAGE[code] : void 0;
+}
+function beatSeconds(line, videoModel) {
   const words2 = line.trim().split(/\s+/).length;
   const spoken = words2 / 3.06;
-  const allowed = [...SEEDANCE_DURATIONS].sort((a, b) => a - b);
+  const allowed = (videoModel ? clipDurationsFor(videoModel) : void 0) ?? [...SEEDANCE_DURATIONS].sort((a, b) => a - b);
   return allowed.find((d) => d >= spoken) ?? allowed[allowed.length - 1];
 }
 var BRAND_PLATE = "A solid flat brand colour plate \u2014 an end card. No people, no objects, no scenery, no text of any kind.";
@@ -38553,18 +38659,21 @@ function castedAvatarOnRoster(ad, roster) {
   if (ad.cast?.avatar) return null;
   const haystack = [ad.cast?.description ?? "", ...(ad.beats ?? []).map((b) => b.show ?? "")].join(" ");
   if (!haystack.trim()) return null;
-  const words2 = new Set(
-    haystack.toLowerCase().split(/[^\p{L}\p{N}]+/u).filter(Boolean)
-  );
+  const written = new Set(haystack.split(/[^\p{L}\p{N}]+/u).filter(Boolean));
   for (const person of roster) {
-    for (const label of [person.name, person.handle]) {
-      const token = label.trim().toLowerCase();
-      if (!token || token.includes("-") || token.length < 3) continue;
-      if (words2.has(token)) return person.handle;
-    }
+    if (labelWrittenAsName(person.name, written) || labelWrittenAsName(person.handle, written)) return person.handle;
   }
   return null;
 }
+function labelWrittenAsName(label, written) {
+  const token = label.trim();
+  if (!token || token.includes("-") || token.length < 3) return false;
+  const folded = token.toLowerCase();
+  for (const word of written) {
+    if (word.toLowerCase() === folded && word[0] === word[0]?.toUpperCase()) return true;
+  }
+  return false;
+}
 function adSpecCastElements(spec, avatar) {
   const lastIndex = spec.beats.length - 1;
   const speakingScenes = spec.beats.map((_, i) => i).filter((i) => !(endCardWanted(spec) && i === lastIndex));
@@ -38606,11 +38715,11 @@ function motionPrompt(beat, toCamera, isClosingCard, place, physics) {
   if (toCamera) return shot;
   return `${shot} Nobody in frame is speaking: mouths CLOSED and STILL throughout, no talking, no lip movement. Whoever is in shot is looking, reacting or working \u2014 never addressing the camera.`;
 }
-function staticTranscript(spec) {
+function staticTranscript(spec, videoModel) {
   const out = [];
   let clock = 0;
   for (const beat of spec.beats) {
-    const duration = beatSeconds(beat.say);
+    const duration = beatSeconds(beat.say, videoModel);
     const words2 = beat.say.trim().split(/\s+/).filter(Boolean);
     const each = words2.length > 0 ? duration / words2.length : duration;
     words2.forEach((text2, i) => {
@@ -38676,9 +38785,12 @@ function voiceoverMode(spec, nativeSpeech) {
 function castRole(spec) {
   if (!spec.cast) return void 0;
   const described = spec.cast.description?.trim();
-  return described ? `${described} \u2014 a customer telling their own story, not a worker` : "a customer telling their own story, not a worker";
+  if (!described) return "a customer telling their own story, not a worker";
+  return NAMES_A_JOB.test(described) ? described : `${described} \u2014 a customer telling their own story, not a worker`;
 }
-function adSpecToBlueprint(spec, avatarAccent) {
+var NAMES_A_JOB = /\b(installer|instalador[ao]?|técnic[oa]|tecnic[oa]|technician|engineer|ingenier[oa]|employee|emplead[oa]|worker|obrer[oa]|staff|clinician|doctor|médic[oa]|medic[oa]|nurse|enfermer[oa]|chef|cocioner[oa]|cociner[oa]|barista|stylist|estilista|trainer|entrenador[ao]?|founder|fundador[ao]?|owner|dueñ[oa]|our team|nuestro equipo|on site|en obra)\b/i;
+function adSpecToBlueprint(spec, opts = {}) {
+  const avatarAccent = opts.avatarAccent;
   const total = spec.beats.length;
   const market = marketFor(spec);
   const place = market ? ` Set in ${market}: the architecture, streets and styling are ${market}'s.` : "";
@@ -38689,7 +38801,7 @@ function adSpecToBlueprint(spec, avatarAccent) {
   let clock = 0;
   const scenes = spec.beats.map((beat, i) => {
     const isClosingCard = closesOnCard && i === total - 1;
-    const duration = beatSeconds(beat.say);
+    const duration = beatSeconds(beat.say, opts.videoModel);
     const start = clock;
     clock += duration;
     return {
@@ -39160,6 +39272,18 @@ var scaffoldAdCommand = defineCommand111({
         avatarError = `Could not read avatar \`${handle}\`: ${e instanceof Error ? e.message : String(e)}. \`baker avatars list\` shows who there is.`;
       }
     }
+    if (avatarError && !spec.data.cast?.description?.trim()) {
+      writeJson({
+        ok: false,
+        error: {
+          code: "VALIDATION_ERROR",
+          message: avatarError,
+          fix: `This ad casts \`${handle}\` and describes nobody else, so without them every shot renders a person nobody chose. Wait for the avatar and re-run \u2014 \`baker avatars get ${handle} --wait\` returns the moment it is ready \u2014 or cast someone who is (\`baker avatars list\`). To build it with an invented person instead, put a written description in \`cast.description\`.`
+        }
+      });
+      process.exit(1);
+      return;
+    }
     const fromWorkspace = await readBrandFromWorkspace();
     const filledLogo = !spec.data.brand?.logo && fromWorkspace.logo ? fromWorkspace.logo : void 0;
     const filledPalette = !spec.data.brand?.palette?.length && fromWorkspace.palette?.length ? fromWorkspace.palette : void 0;
@@ -39238,7 +39362,7 @@ var scaffoldAdCommand = defineCommand111({
       process.exit(1);
       return;
     }
-    const blueprint = adSpecToBlueprint(spec.data, avatar?.profile?.accent);
+    const blueprint = adSpecToBlueprint(spec.data, { avatarAccent: avatar?.profile?.accent, videoModel: chosenVideoModel });
     const slug = args.slug ?? path24.basename(specPath).replace(/\.[^.]+$/, "");
     const outPath = args.out ?? (args.slug ? path24.join("src/creatives", slug, `${slug}.canvas.json`) : path24.join(path24.dirname(specPath), `${slug}.canvas.json`));
     const outDir = path24.dirname(outPath);
@@ -39251,10 +39375,10 @@ var scaffoldAdCommand = defineCommand111({
     const blueprintStylePath = path24.join(outDir, "prompt.style.json");
     await writeSceneFiles(outDir, blueprint);
     await writeFile9(blueprintStylePath, renderStyleProjectionFromValue(blueprint), "utf8");
-    const logoPath = spec.data.brand?.logo?.trim();
+    const logoPath = noBrand ? void 0 : spec.data.brand?.logo?.trim();
     const transcriptPath = path24.join(outDir, "transcript.json");
     if (spec.data.voiceover === false) {
-      await writeFile9(transcriptPath, `${JSON.stringify(staticTranscript(spec.data), null, 2)}
+      await writeFile9(transcriptPath, `${JSON.stringify(staticTranscript(spec.data, chosenVideoModel), null, 2)}
 `, "utf-8");
     }
     const opts = {
@@ -39283,6 +39407,12 @@ var scaffoldAdCommand = defineCommand111({
       // its accent. The clip speaks in the voice the video model renders, and the region is
       // directed in the prompt, which is the only place a pronunciation can come from.
       ...avatar?.profile?.accent?.trim() ? { spokenAccent: avatar.profile.accent.trim() } : {},
+      // Same decision, one step further: with an avatar on camera the clip's own voice is
+      // kept, not re-voiced. The re-voice is the speech-to-speech route ADR 0005 measured
+      // and rejected — it would swap the voice that matches the mouth and the directed
+      // region for one cast off an English stock list. Only the on-camera voice; a
+      // voiceover read has no rendered voice to keep and is still cast.
+      ...avatar ? { keepRenderedVoice: true } : {},
       ...spec.data.voiceover === false ? { staticTranscriptPath: path24.relative(outDir, transcriptPath) } : {}
     };
     const canvas = scaffoldVideoCanvas(blueprint, adSpecCastElements(spec.data, avatar), opts);
@@ -39298,7 +39428,7 @@ var scaffoldAdCommand = defineCommand111({
       }
     }
     const closing = blueprint.scenes[blueprint.scenes.length - 1];
-    const overlayHtml = endCardWanted(spec.data) && closing && renderedDir ? buildAdBrandOverlay({
+    const overlayHtml = !noBrand && endCardWanted(spec.data) && closing && renderedDir ? buildAdBrandOverlay({
       logo: staged,
       cta: endCardCta(spec.data),
       cardAtS: closing.start_s,
@@ -50871,6 +51001,20 @@ function blankAstroFrontmatter(text2) {
   }
   return kept.join("") + text2.slice(cut);
 }
+function blankUnrenderedSource(text2) {
+  return blankTemplateComments(blankAstroFrontmatter(text2));
+}
+function blankTemplateComments(text2) {
+  const out = text2.split("");
+  const pattern = /\{\s*\/\*[\s\S]*?\*\/\s*\}|<!--[\s\S]*?-->/g;
+  for (const m of text2.matchAll(pattern)) {
+    const start = m.index ?? 0;
+    for (let i = start; i < start + m[0].length && i < out.length; i++) {
+      if (out[i] !== "\n") out[i] = " ";
+    }
+  }
+  return out.join("");
+}
 function frontmatterLength(text2) {
   if (!text2.startsWith("---")) return 0;
   const lines = text2.split("\n");
@@ -51244,7 +51388,7 @@ function distinctCompetitorNames(competitors) {
   return [...byFold.values()];
 }
 function mentionsInSource(source, names) {
-  const text2 = blankAstroFrontmatter(source.text);
+  const text2 = blankUnrenderedSource(source.text);
   const visible = blankHtmlKeepingOffsets(text2);
   const tables = tableRanges(text2);
   const out = [];
@@ -51284,7 +51428,7 @@ function escapeRegExp(value) {
 }
 function detectSource(source) {
   const file = source.path;
-  const text2 = blankAstroFrontmatter(source.text);
+  const text2 = blankUnrenderedSource(source.text);
   const findings = [];
   const lines = text2.split("\n");
   for (const matcher of LINE_MATCHERS) {
@@ -51302,7 +51446,7 @@ function detectSource(source) {
 }
 function detectPage(sources) {
   if (sources.length === 0) return [];
-  const combined = sources.map((s) => blankAstroFrontmatter(s.text)).join("\n");
+  const combined = sources.map((s) => blankUnrenderedSource(s.text)).join("\n");
   const pageFile = sources.find((s) => s.path === "index.astro" || s.path.endsWith("/index.astro"))?.path ?? sources[0]?.path ?? "index.astro";
   const findings = [];
   for (const analyzer of ANALYZERS) {
@@ -51382,7 +51526,7 @@ function describeCounts(findings) {
 // src/commands/landing/snapshot.ts
 import { mkdir as mkdir10, rename as rename2, writeFile as writeFile14 } from "fs/promises";
 import path34 from "path";
-var CRITIC_VERSION = "3";
+var CRITIC_VERSION = "4";
 function critiqueCacheDir(projectRoot) {
   return path34.join(projectRoot, ".cache", "landing-critique");
 }
@@ -57902,6 +58046,8 @@ function costHintsFor(body) {
   if (pinned) hints2.push(pinned);
   const pace = spokenPaceHint(body.prompt, body.duration, model);
   if (pace) hints2.push(pace);
+  const silent = silentModelHint(model, body.prompt);
+  if (silent) hints2.push(silent);
   if (count > 1) {
     hints2.push(
       "Ask for more than one take only when the MOTION is the risk \u2014 for a look you are unsure about, iterate on the still first."