@ossclip/core 0.1.24 → 0.1.25

package/src/producer/youtube.ts

@@ -0,0 +1,434 @@
+import { z } from "zod/v4";
+import type { Word } from "../schema";
+import type { TimeMap } from "../timemap";
+import { isSentenceEnd } from "../clip";
+import type { LlmProvider } from "./provider";
+import { cappedText } from "./beats";
+
+/**
+ * The `--youtube` SEO pack (Y2, 2026-08-16; prompt v2 2026-08-17): title
+ * options, a description with measured chapter timestamps spliced in,
+ * hashtags and tags, written beside the video as `<out>.youtube.md` so the
+ * upload form is a paste job instead of a second writing session.
+ *
+ * Free text is `cappedText`, not `.max()` — the §112 doctrine as applied in
+ * beats.ts: LLM output is untrusted input, validated where the pipeline can
+ * still degrade instead of at the point where it can only die. A title one
+ * word over budget must cost a word, never the run.
+ *
+ * Two sections a human YouTube-strategist workflow would include are
+ * deliberately NOT here (prompt v2 decision):
+ *  - competitor analysis — the provider has no browsing, so "competitor
+ *    research" would be confident hallucination dressed up as strategy;
+ *  - a thumbnail prompt — ossclip GENERATES the actual thumbnail from the
+ *    real portrait (thumbnail.ts), which beats prose describing an imaginary
+ *    one.
+ */
+
+/**
+ * Bump whenever buildYoutubePrompt changes what it asks for: a prompt change
+ * changes the answer, so the Y2 pack cache key carries this (the §78
+ * cache-key posture) — an old cached pack must not survive a new prompt.
+ */
+export const YOUTUBE_PROMPT_VERSION = "v2";
+
+export const YoutubeChapterSchema = z.object({
+  /** Output-timeline seconds — the produced video's clock, not the source's. */
+  atSec: z.number().nonnegative(),
+  title: cappedText(80),
+});
+export type YoutubeChapter = z.infer<typeof YoutubeChapterSchema>;
+
+/**
+ * The three labeled title angles for YouTube's Test & Compare A/B feature
+ * (prompt v2): browse = story/curiosity for home-page CTR, search =
+ * keyword/tool-name for evergreen search, benefit = the ROI claim. "alt"
+ * covers any extra option past the three.
+ */
+export const TitleAngleSchema = z.enum(["browse", "search", "benefit", "alt"]);
+export type TitleAngle = z.infer<typeof TitleAngleSchema>;
+
+export const YoutubePackSchema = z.object({
+  /** Distinct angles on the same video, not five rewordings of one. */
+  titles: z.array(cappedText(100)).min(3).max(5),
+  /**
+   * Parallel to `titles`, one angle label each. OPTIONAL for back-compat:
+   * packs approved before prompt v2 carry bare titles, and the approved-file
+   * contract (YOUTUBE_APPROVED_BASENAME) means those files must keep parsing
+   * verbatim forever.
+   */
+  titleAngles: z.array(TitleAngleSchema).optional(),
+  /** Prompted for keyword-first-lines, hashtags-at-end; free-form otherwise.
+   * Chapter timestamps are NOT the model's to write — the formatter splices
+   * the measured ones in (spliceTimestampsIntoDescription). */
+  description: z.string(),
+  hashtags: z.array(z.string()),
+  tags: z.array(z.string()),
+  chapters: z.array(YoutubeChapterSchema).optional(),
+  /** Advisory first-60-seconds retention note for the CREATOR — never pasted
+   * into YouTube. Optional: pre-v2 packs never carried it. */
+  hook60: cappedText(400).optional(),
+  /** A ready-to-post LinkedIn announcement. Optional, pre-v2 back-compat. */
+  linkedinPost: cappedText(1500).optional(),
+  /** A short YouTube community post for existing subscribers. Optional. */
+  communityPost: cappedText(400).optional(),
+});
+export type YoutubePack = z.infer<typeof YoutubePackSchema>;
+
+/** YouTube's hard cap on the tags field, counted over the comma-joined text. */
+export const YOUTUBE_TAGS_LIMIT = 500;
+
+/**
+ * Post-parse guard for the tags field: drop tags from the END until the
+ * comma-joined line fits YouTube's limit. From the end because the prompt
+ * asks for relevance order — the least relevant tag is the one a budget cut
+ * should cost. Degrade-don't-die (§112): the schema cannot express a
+ * joined-length cap, and refusing the whole pack over an over-enthusiastic
+ * tag list would discard the titles and description that were fine.
+ */
+export function trimTagsToLimit(tags: string[], limit: number = YOUTUBE_TAGS_LIMIT): string[] {
+  const kept = [...tags];
+  while (kept.length > 0 && kept.join(", ").length > limit) {
+    kept.pop();
+  }
+  return kept;
+}
+
+/**
+ * The workdir file the editor's SEO panel writes (2026-08-17), mirroring the
+ * thumbnail approval contract (THUMBNAIL_APPROVED_BASENAME, thumbnail.ts):
+ * once it exists, produce's Y2 block uses the pack VERBATIM — no cache
+ * lookup, no LLM call — because an edited pack is the user's decision, and a
+ * fresh generation would silently discard it. Holds a plain `YoutubePack`,
+ * no skip variant: the `--youtube` flag itself gates the feature, so there
+ * is no "opted in then declined" state for a pack the way there is for a
+ * thumbnail. Deleting the file is the regenerate gesture.
+ */
+export const YOUTUBE_APPROVED_BASENAME = "youtube-pack-approved.json";
+
+/**
+ * How much transcript the prompt carries. Field-measured 2026-08-17: at the
+ * old 10k cap a 21-minute video's chapters STOPPED AT 8:32 — the model can
+ * only chapter what it reads, and a truncated transcript silently produces
+ * chapters for the excerpt while looking complete. 60k chars ≈ 15k tokens
+ * covers ~45 minutes of stamped talk for pennies; past that the truncation
+ * note at least tells the model (and the chapters) where the coverage ends.
+ * `stampedTranscript` pre-caps to this same constant, so
+ * `buildYoutubePrompt`'s own cap is a backstop for raw-text callers, never a
+ * second cut of stamped input.
+ */
+export const YOUTUBE_TRANSCRIPT_CHAR_CAP = 60_000;
+
+/** 125 → "2:05" — YouTube's own chapter-list spelling. */
+function chapterStamp(sec: number): string {
+  const whole = Math.max(0, Math.floor(sec));
+  const m = Math.floor(whole / 60);
+  const s = whole % 60;
+  return `${m}:${String(s).padStart(2, "0")}`;
+}
+
+/** What a truncated stamped transcript ends with — the model must know it is
+ * reading an excerpt, or it will write a description that promises only the
+ * first half. */
+const TRUNCATION_NOTE = "[transcript truncated — the video continues]";
+
+/**
+ * The transcript as the prompt's chapter evidence (prompt v2): one line per
+ * sentence, prefixed with the sentence's first surviving word's OUTPUT time
+ * as `[m:ss]`. This is the pack's unique advantage over paste-a-transcript
+ * prompt tools — words carry SOURCE times and the TimeMap converts them to
+ * the produced video's clock, so chapters can be measured instead of the
+ * guessed "0:00 Intro" everyone else ships.
+ *
+ * Fully-cut sentences (every word `mapWord` → null) are skipped: they are
+ * not in the video, and stamping them would hand the model timestamps that
+ * point at content the viewer never sees. Pure — words + map in, string
+ * out — so the mapping, the cut-skip and the cap are testable without an
+ * LLM or a filesystem.
+ *
+ * The cap reserves room for TRUNCATION_NOTE up front, so the result NEVER
+ * exceeds `maxChars` — truncated or not — and whole sentences are dropped
+ * rather than sliced mid-line, which would leave a dangling stamp.
+ */
+export function stampedTranscript(
+  words: readonly Word[],
+  map: TimeMap,
+  opts: { maxChars?: number } = {},
+): string {
+  const maxChars = opts.maxChars ?? YOUTUBE_TRANSCRIPT_CHAR_CAP;
+  // isSentenceEnd reads punctuation off Transcript-shaped input.
+  const t = { language: "en", words: [...words] };
+  const budget = maxChars - TRUNCATION_NOTE.length - 1;
+  const lines: string[] = [];
+  let used = 0;
+  let truncated = false;
+  let start = 0;
+  for (let i = 0; i < words.length; i++) {
+    if (i < words.length - 1 && !isSentenceEnd(t, i)) continue;
+    const sentence = words.slice(start, i + 1);
+    start = i + 1;
+    // The stamp is the first SURVIVING word's output time, not the first
+    // word's: a sentence whose opening filler was cut starts, on the
+    // viewer's clock, at the first word they hear.
+    let stampSec: number | null = null;
+    for (const w of sentence) {
+      const mapped = map.mapWord(w);
+      if (mapped) {
+        stampSec = mapped.start;
+        break;
+      }
+    }
+    if (stampSec === null) continue; // the whole sentence was cut
+    const line = `[${chapterStamp(stampSec)}] ${sentence.map((w) => w.text).join(" ")}`;
+    const cost = lines.length === 0 ? line.length : line.length + 1;
+    if (used + cost > budget) {
+      truncated = true;
+      break;
+    }
+    lines.push(line);
+    used += cost;
+  }
+  if (truncated) lines.push(TRUNCATION_NOTE);
+  return lines.join("\n");
+}
+
+/** YouTube ignores a chapter list whose entries run shorter than this. */
+export const YOUTUBE_CHAPTER_MIN_GAP_SEC = 10;
+/** ...or that carries fewer timestamps than this. */
+export const YOUTUBE_CHAPTER_MIN_COUNT = 3;
+
+/**
+ * Post-parse guard for the chapters field — trimTagsToLimit's posture
+ * applied to YouTube's chapter rules, which the schema cannot express:
+ * the list must start at 0:00, hold ≥3 entries, and give every chapter
+ * ≥10 seconds, or YouTube silently ignores the WHOLE list. Degrade-don't-die
+ * (§112): repair what can be repaired, drop what cannot, and return [] —
+ * "no chapters", which the formatter renders as no timestamp block — rather
+ * than ship a list YouTube would reject wholesale anyway.
+ *
+ *  - unsorted input sorts ascending (the model's list order is untrusted);
+ *  - chapters at or past the video's end are dropped — a stamp the video
+ *    never reaches is a hallucinated one;
+ *  - a first chapter within the 10s minimum of 0 snaps TO 0 (it IS the
+ *    intro, mis-stamped); one further out gets a synthetic 0:00 "Intro"
+ *    prepended — the unnamed opening is real content, and 0:00 is mandatory;
+ *  - of a pair closer than 10s, the LATER one drops (the earlier stamp is
+ *    the section's true start).
+ */
+export function normalizeChapters(
+  chapters: readonly YoutubeChapter[],
+  durationSec: number,
+): YoutubeChapter[] {
+  const sorted = [...chapters]
+    .sort((a, b) => a.atSec - b.atSec)
+    .filter((c) => c.atSec < durationSec);
+  if (sorted.length === 0) return [];
+  const first = sorted[0]!;
+  const list: YoutubeChapter[] =
+    first.atSec === 0
+      ? sorted
+      : first.atSec < YOUTUBE_CHAPTER_MIN_GAP_SEC
+        ? [{ ...first, atSec: 0 }, ...sorted.slice(1)]
+        : [{ atSec: 0, title: "Intro" }, ...sorted];
+  const kept: YoutubeChapter[] = [];
+  for (const c of list) {
+    const prev = kept[kept.length - 1];
+    if (prev === undefined || c.atSec - prev.atSec >= YOUTUBE_CHAPTER_MIN_GAP_SEC) kept.push(c);
+  }
+  return kept.length >= YOUTUBE_CHAPTER_MIN_COUNT ? kept : [];
+}
+
+export interface YoutubePromptArgs {
+  /** The stamped transcript (stampedTranscript) — sentences prefixed with
+   * their `[m:ss]` output-clock time. Plain text still works (raw-text
+   * callers just get no measured chapters worth trusting). */
+  transcriptText: string;
+  /** `--intent`, when the run had one. */
+  intent?: string;
+  /** The producer's hook, when a beat sheet exists — the strongest claim. */
+  hook?: string;
+  /** The cover banner text, when a beat sheet wrote one. */
+  coverText?: string;
+  /**
+   * Who watches the channel (`--audience` / config `audience`, 2026-08-16):
+   * titles and tags for "junior devs learning agents" are a different pack
+   * than for "engineering managers", and only the user knows which one this
+   * channel is.
+   */
+  audience?: string;
+  /** Output duration in seconds — the ceiling no chapter may pass. */
+  durationSec: number;
+}
+
+/**
+ * Pure prompt builder, separated from the provider call so the
+ * include/omit matrix (intent, hook, coverText, audience) and the transcript
+ * cap are testable without an LLM.
+ */
+export function buildYoutubePrompt(args: YoutubePromptArgs): { system: string; user: string } {
+  const system =
+    "You are an expert YouTube growth strategist writing upload metadata for a finished video. " +
+    "The transcript's sentences are each prefixed with a [m:ss] timestamp on the FINAL video's " +
+    "clock — measured from the actual edit, not estimated. Match the creator's tone from the " +
+    "transcript, and never make a claim the video does not deliver: click-drivers that deliver, " +
+    "not clickbait.\n" +
+    "- titles: 3-5 options for YouTube's Test & Compare A/B feature, each a DISTINCT angle, " +
+    "labeled in the parallel titleAngles array: \"browse\" — story/curiosity, zero jargon, " +
+    "before/after tension, built for home-page CTR; \"search\" — keyword-heavy, names the exact " +
+    "tools and technologies, built for evergreen search; \"benefit\" — the ROI, \"do X in N " +
+    "minutes\". Include at least one of each of those three; label any extra option \"alt\". " +
+    "Target at most 70 characters (search results truncate around 70, browse cards earlier).\n" +
+    "- description: the FIRST TWO lines are the search snippet — the primary keyword plus a " +
+    "curiosity hook. Then a body of short paragraphs or bullets carrying the secondary keywords, " +
+    "a call to action near the end, and 3-5 hashtags as the LAST line. Do NOT write timestamp " +
+    "or chapter lines in the description — the measured chapter timestamps are spliced in " +
+    "automatically.\n" +
+    "- hashtags: the same 3-5, as an array.\n" +
+    "- tags: search keywords and phrases a viewer would type, ordered by relevance — the most " +
+    "important first.\n" +
+    "- chapters: 4-10, USING ONLY the [m:ss] stamps that appear in the transcript — atSec is " +
+    "that stamp converted to seconds, never a time you estimated. The first chapter is at 0. " +
+    "Titles at most 50 characters, keyword-bearing but written for a human scanning the list.\n" +
+    "- hook60: a first-60-seconds retention note for the CREATOR (it is never published) — e.g. " +
+    "\"show the final result at 0:05, then rewind to how\" — grounded in what this transcript " +
+    "actually opens with.\n" +
+    "- linkedinPost: a story-driven LinkedIn post about this video: short lines with line " +
+    "breaks, a curiosity gap, no hashtag spam, ending by pointing to the link in the comments " +
+    "(the LinkedIn convention for off-platform links).\n" +
+    "- communityPost: a short, casual YouTube community post for existing subscribers.";
+  const capped =
+    args.transcriptText.length > YOUTUBE_TRANSCRIPT_CHAR_CAP
+      ? // Slice + say so: the model must know it is reading an excerpt, or it
+        // will write a description that promises only the first half.
+        `${args.transcriptText.slice(0, YOUTUBE_TRANSCRIPT_CHAR_CAP)}\n${TRUNCATION_NOTE}`
+      : args.transcriptText;
+  const user =
+    (args.intent ? `Intent: ${args.intent}\n` : "") +
+    (args.hook ? `Hook (already chosen by the producer): ${args.hook}\n` : "") +
+    (args.coverText ? `Cover headline: ${args.coverText}\n` : "") +
+    (args.audience ? `Audience: ${args.audience}\n` : "") +
+    `Total runtime: ${chapterStamp(args.durationSec)} — chapters must not exceed it.\n\n` +
+    `Transcript:\n${capped}`;
+  return { system, user };
+}
+
+/** One editorial call → a validated, tag-trimmed, chapter-normalized pack. */
+export async function generateYoutubePack(
+  provider: LlmProvider,
+  args: YoutubePromptArgs,
+): Promise<YoutubePack> {
+  const { system, user } = buildYoutubePrompt(args);
+  const pack = await provider.complete({
+    system,
+    user,
+    schema: YoutubePackSchema,
+    schemaName: "youtube_pack",
+    tier: "editorial",
+  });
+  // Both post-parse guards live HERE, beside the call, so every path that
+  // stores the pack (the Y2 cache, the approved file) already holds a
+  // publishable one. An empty normalized list degrades to "no chapters"
+  // (undefined), which the formatter renders as no timestamp block.
+  const chapters = normalizeChapters(pack.chapters ?? [], args.durationSec);
+  return {
+    ...pack,
+    tags: trimTagsToLimit(pack.tags),
+    chapters: chapters.length > 0 ? chapters : undefined,
+  };
+}
+
+/** The heading the timestamp block opens with inside the description. */
+export const YOUTUBE_TIMESTAMPS_HEADING = "⏱️ Timestamps:";
+
+/** A line that is nothing but hashtags — the description's prompted last
+ * line, which the timestamp splice must stay above. */
+function isHashtagLine(line: string): boolean {
+  const tokens = line.trim().split(/\s+/);
+  return tokens.length > 0 && tokens.every((t) => t.startsWith("#") && t.length > 1);
+}
+
+/**
+ * Splice the measured chapter list INTO the description text — YouTube
+ * parses chapters FROM the description, so a separate list beside it would
+ * make the user hand-merge two blocks the file exists to spare them. The
+ * block lands after the body but ABOVE a trailing hashtag line when the
+ * description has one (the prompt asks for hashtags last, and search treats
+ * that placement specially). Pure and exported so the above/below matrix is
+ * testable on its own.
+ */
+export function spliceTimestampsIntoDescription(
+  description: string,
+  chapters: readonly YoutubeChapter[],
+): string {
+  const body = description.trimEnd();
+  if (chapters.length === 0) return body;
+  const block = [
+    YOUTUBE_TIMESTAMPS_HEADING,
+    ...chapters.map((c) => `${chapterStamp(c.atSec)} ${c.title}`),
+  ].join("\n");
+  const lines = body.split("\n");
+  const last = lines[lines.length - 1] ?? "";
+  if (lines.length > 1 && isHashtagLine(last)) {
+    const head = lines.slice(0, -1).join("\n").trimEnd();
+    return `${head}\n\n${block}\n\n${last}`;
+  }
+  return `${body}\n\n${block}`;
+}
+
+/** How a titleAngle spells inside the markdown's numbered list. */
+const ANGLE_LABELS: Record<TitleAngle, string> = {
+  browse: "Browse",
+  search: "Search",
+  benefit: "Benefit",
+  alt: "Alt",
+};
+
+/**
+ * The pack as the markdown file a user pastes from. Pure — the produce
+ * orchestration owns the write.
+ *
+ * Chapters render INSIDE the description (spliceTimestampsIntoDescription),
+ * not as their own section: the description is the one paste target, and a
+ * separate `## Chapters` block was a second copy of the same lines. The
+ * optional prompt-v2 sections (hook strategy, LinkedIn, community) render
+ * only when the pack carries them, so a pre-v2 pack formats exactly as it
+ * always did.
+ *
+ * `chaptersFromSec` rebases chapter stamps by subtracting it (dropping any
+ * chapter that would land before 0), for a caller whose chapters are on some
+ * other clock than the output's. Default 0 — stamps pass through.
+ */
+export function formatYoutubeMarkdown(
+  pack: YoutubePack,
+  opts: { chaptersFromSec?: number } = {},
+): string {
+  const lines: string[] = ["# YouTube pack", "", "## Title options", ""];
+  pack.titles.forEach((t, i) => {
+    // Labels only when the pack carries angles (prompt v2); a mismatched or
+    // missing entry renders the bare title — the label is a hint, not data
+    // the paste depends on.
+    const angle = pack.titleAngles?.[i];
+    lines.push(`${i + 1}. ${angle !== undefined ? `[${ANGLE_LABELS[angle]}] ` : ""}${t}`);
+  });
+  const from = opts.chaptersFromSec ?? 0;
+  const chapters = (pack.chapters ?? [])
+    .filter((c) => c.atSec - from >= 0)
+    .map((c) => ({ ...c, atSec: c.atSec - from }));
+  lines.push(
+    "",
+    "## Description",
+    "",
+    spliceTimestampsIntoDescription(pack.description, chapters),
+    "",
+    "## Hashtags",
+    "",
+  );
+  // Paste-ready: one line, every entry carrying its # exactly once whether or
+  // not the model already spelled it.
+  lines.push(pack.hashtags.map((h) => (h.startsWith("#") ? h : `#${h}`)).join(" "));
+  lines.push("", "## Tags (comma-separated)", "", pack.tags.join(", "));
+  if (pack.hook60) lines.push("", "## First-60s hook strategy", "", pack.hook60.trimEnd());
+  if (pack.linkedinPost) lines.push("", "## LinkedIn post", "", pack.linkedinPost.trimEnd());
+  if (pack.communityPost) lines.push("", "## Community post", "", pack.communityPost.trimEnd());
+  return `${lines.join("\n")}\n`;
+}

package/src/retake.ts

@@ -20,6 +20,21 @@ import type { Analysis, Span, Transcript } from "./schema";
  * (edit distance over the token stream, not the letters) rather than a single
  * fuzzy-string threshold: it tolerates the fuzzy word or two `soundsSimilar`
  * chased, without opening the same phonetic false-positive channel.
+ *
+ * The exact-prefix restart pass (2026-08-16 field case) deliberately NARROWS
+ * a §128 protection rather than repealing it. §128 (PHASE1-FINDINGS.md,
+ * "128. Retake collapse without a spoken marker") records the verified repro
+ * where "Let me show you this." scored a spurious 1.0 against its own longer
+ * continuation and the LONGER, more complete side got cut — which is why
+ * complete-vs-complete comparison is full-sequence and a shorter complete
+ * sentence never prefix-matches forward. The field flub this pass exists for
+ * is the opposite shape: "You can use OpenAI." → [pause] → "You can use
+ * OpenAI models and whatnot." — ASR punctuated the abandoned attempt as
+ * complete, so the protection above (correctly) refuses the match, and the
+ * flub ships. The pass cuts ONLY the strictly-shorter EARLIER sentence, only
+ * on an EXACT token prefix, and only with a real pause at the boundary —
+ * three conditions the §128 incident's cut (of the longer side, on a fuzzy
+ * score, with no pause requirement) would fail.
  */
 
 /** Sequence similarity floor for two attempts to count as the same line. */
@@ -58,6 +73,25 @@ export const HALLUCINATION_SILENCE_FRAC = 0.65;
  * be at this boundary, so the same threshold gates both: not tuned twice.
  */
 export const RESTART_SPLIT_MIN_SIL = 0.35;
+/**
+ * How far INTO the next sentence's stamped words the exact-prefix pass looks
+ * for the restart pause. Stamp-stretch physics again (§18): the pause between
+ * the abandoned "You can use OpenAI." and its restart was the silence
+ * 201.52–201.93 — INSIDE the stamped span of the restart sentence
+ * (201.0–203.08), not in the inter-sentence gap, because whisper stretched
+ * the restart's opening words back over the dead air. So the window is
+ * [A.endSec, B.startSec + this], measured with `silenceOverlap` against real
+ * silence spans, never against word-stamp gaps (which are ~always zero).
+ */
+export const PREFIX_RESTART_SIL_WINDOW = 1.0;
+/**
+ * Confidence `produce.ts` assigns an exact-prefix restart cut in the
+ * cutlist: below the 0.9 an ordinary similarity-matched retake earns,
+ * because the evidence is structural (prefix + pause) rather than a scored
+ * near-identical repeat — the report and any confidence-gated behavior can
+ * tell the two rules apart.
+ */
+export const RESTART_PREFIX_CONFIDENCE = 0.85;
 
 /** A span of transcript, in word indices (inclusive) and source seconds. */
 export interface RetakeInstance {
@@ -119,6 +153,13 @@ export interface RetakeGroup {
   cuts: RetakeCut[];
   hallucinated: RetakeHallucination[];
   undecided: RetakeUndecided[];
+  /**
+   * Which pass decided this group, when it was not the ordinary similarity
+   * chain: "exact-prefix" is the restart pass (module doc) — the report
+   * prints the structural evidence instead of a match percentage, because
+   * "100% match" would misdescribe a rule that never scored anything.
+   */
+  rule?: "exact-prefix";
 }
 
 // ---- token comparison -------------------------------------------------
@@ -577,9 +618,63 @@ export function findRetakeGroups(
     }
     if (overlaps) continue;
     groups.push(g);
+    // The sentence pass's members join the claimed set too: the exact-prefix
+    // pass below must not re-decide words either earlier pass already
+    // decided — including spares, so a declined cut stays declined.
+    for (const m of members) {
+      for (let w = m.startWord; w <= m.endWord; w++) claimed.add(w);
+    }
+  }
+
+  // Exact-prefix restart pass (2026-08-16 field case; module doc has the
+  // §128-narrowing rationale). The flub: sentence A is an EXACT token prefix
+  // of the immediately following sentence B, strictly shorter, with a real
+  // pause (>= RESTART_SPLIT_MIN_SIL of silence) at the restart boundary. ASR
+  // punctuated A as complete ("You can use OpenAI."), so the similarity
+  // passes above score A-vs-B honestly divergent (1 - 3/7 ≈ 0.57) and
+  // correctly decline — completeness is no defense here, which is why A may
+  // be complete. Adjacency skips zero-token sentences, the same
+  // filler/marker transparency the chaining rule grants.
+  const live = sentences.filter((s) => s.tokens.length > 0);
+  for (let k = 0; k + 1 < live.length; k++) {
+    const a = live[k]!;
+    const b = live[k + 1]!;
+    if (a.hallucinated || b.hallucinated) continue;
+    if (a.tokens.length < RETAKE_MIN_TOKENS) continue;
+    if (a.tokens.length >= b.tokens.length) continue;
+    if (!a.tokens.every((tok, j) => tokensEqual(tok, b.tokens[j]!))) continue;
+    // The pause: ONE silence overlapping [A.end, B.start + window] by at
+    // least RESTART_SPLIT_MIN_SIL — see PREFIX_RESTART_SIL_WINDOW for why
+    // the window reaches INTO B's stamped words and why word gaps are
+    // useless here. Per-silence, not summed: three 0.1s breath gaps across
+    // the window are ordinary delivery, not a restart pause.
+    const windowEnd = b.startSec + PREFIX_RESTART_SIL_WINDOW;
+    const paused = analysis.silences.some(
+      (s) => silenceOverlap([s], a.endSec, windowEnd) >= RESTART_SPLIT_MIN_SIL,
+    );
+    if (!paused) continue;
+    // Claimed dedupe, same set as the sentence pass: a word either earlier
+    // pass already cut, kept OR deliberately spared is not re-decidable.
+    let overlaps = false;
+    for (const m of [a, b]) {
+      for (let w = m.startWord; w <= m.endWord && !overlaps; w++) {
+        if (claimed.has(w)) overlaps = true;
+      }
+    }
+    if (overlaps) continue;
+    groups.push({
+      kept: toPublic(b),
+      cuts: [{ ...toPublic(a), similarity: prefixSimilarity(a.tokens, b.tokens) }],
+      hallucinated: [],
+      undecided: [],
+      rule: "exact-prefix",
+    });
+    for (const m of [a, b]) {
+      for (let w = m.startWord; w <= m.endWord; w++) claimed.add(w);
+    }
   }
 
-  // Two passes can interleave in transcript order — the report reads in
+  // The passes can interleave in transcript order — the report reads in
   // word order, whichever pass found the group.
   groups.sort((a, b) => {
     const first = (g: RetakeGroup): number =>
@@ -667,7 +762,14 @@ export function formatRetakeGroup(transcript: Transcript, group: RetakeGroup): s
   } else {
     lines.push(`kept: "${said(group.kept)}"`);
     for (const c of group.cuts) {
-      lines.push(`  cut (${Math.round(c.similarity * 100)}% match): "${said(c)}"`);
+      // An exact-prefix cut was never scored — printing a percentage would
+      // claim evidence the rule doesn't have. Name the structural evidence
+      // (prefix + pause) instead.
+      lines.push(
+        group.rule === "exact-prefix"
+          ? `  cut (exact-prefix restart, ≥0.35s pause): "${said(c)}"`
+          : `  cut (${Math.round(c.similarity * 100)}% match): "${said(c)}"`,
+      );
     }
     for (const u of group.undecided) {
       const pct = Math.round((u.similarity ?? 0) * 100);

package/src/scene-schema.ts

@@ -186,6 +186,16 @@ export const FaceCropSchema = z.object({
    * phone footage, wrong for a webcam recording or a screen capture.
    */
   sourceAspect: z.number().positive().optional(),
+  /**
+   * What the measured face IS: the frame's subject, or an incidental face in
+   * a frame that is really about something else (2026-08-16 incident: a
+   * screen recording's camera PiP, 12% of frame height at the bottom-right,
+   * dragged the cover crop to the frame bottom and decapitated the speaker
+   * everywhere they appeared full-frame). "screen" tells the stage to center
+   * the cover instead of biasing toward the face. Absent means "face" — every
+   * pre-existing render-props keeps its old behavior.
+   */
+  subject: z.enum(["face", "screen"]).optional(),
 });
 export type FaceCrop = z.infer<typeof FaceCropSchema>;