@banou/media-player 0.10.0 → 0.10.1

package/src/lib/react/video-player.tsx

@@ -1,5 +1,6 @@
 /// <reference types="@emotion/react/types/css-prop" />
 import type { ReactNode } from 'react'
+import type { MediaChapter } from '../engine'
 import type { DownloadedRange } from './source-feature'
 import type { DelegatedTracks, ExternalThumbnails, PlayerMedia } from './media'
 import type { ExposePlayerOptions } from '../remote'
@@ -45,6 +46,19 @@ type CommonOptions = {
   title?: string
   autoplay?: boolean
 
+  /**
+   * Named spans of the timeline, drawn as segments on the seekbar.
+   *
+   * Common to both arms rather than remote-only, unlike `thumbnails`. A local file is just as likely
+   * to declare none (mp4 and webm routinely carry no chapters at all), and a caller often knows
+   * chapters the container does not: skip-intro ranges from a metadata API are the usual case. Given
+   * here they WIN over whatever the container declared, and they paint on the first frame rather than
+   * waiting for the pipeline to boot.
+   *
+   * Expected in seconds, ordered by start and non-overlapping. They need not cover the duration.
+   */
+  chapters?: MediaChapter[]
+
   /**
    * Draw the control bar. Defaults to true.
    *
@@ -183,7 +197,7 @@ const PlayerRoot = ({ options, children }: { options: MediaPlayerOptions, childr
   const local = remote ? null : options as MediaPlayerLocalOptions
   // `title` is common to both arms, so it is read off `options`. Everything else here belongs to the
   // local arm and is absent when the media is remote.
-  const { title } = options
+  const { title, chapters } = options
   const {
     size, downloadedRanges, publicPath, libavWorkerUrl, read, thumbnailRead, thumbnailsEnabled,
   } = local ?? ({} as Partial<MediaPlayerLocalOptions>)
@@ -301,6 +315,19 @@ const PlayerRoot = ({ options, children }: { options: MediaPlayerOptions, childr
     })
   }, [setSourceState, audio])
 
+  /*
+   * Guarded, and its own effect, for the reason the ones above are.
+   *
+   * Folded into the effect that publishes `title` and `size` it would write `chapters: undefined`
+   * every time one of those changed, erasing the list libav had already found. The other half of the
+   * precedence lives at the engine's publish site, which prefers this prop when it has one: between
+   * them the two writers land on the same value whichever runs last.
+   */
+  useEffect(() => {
+    if (!chapters) return
+    setSourceState({ chapters })
+  }, [setSourceState, chapters])
+
   return (
     <Chrome
       ref={setContainer}

package/src/lib/utils/chapters.ts

@@ -0,0 +1,302 @@
+import type { MediaChapter } from '../engine'
+
+/**
+ * Where a chaptered seekbar is broken, and what to paint in each piece.
+ *
+ * Kept out of the component because the geometry is the whole specification of the feature and is
+ * worth testing directly: everything else about chapters on the bar is styling around these two
+ * functions.
+ */
+
+/** Width of the break drawn between two chapters, in px. */
+const CHAPTER_GAP_PX = 2
+/**
+ * A boundary this close to either end of the bar is dropped rather than drawn.
+ *
+ * Containers routinely declare a last chapter ending a few milliseconds before the file does, so
+ * drawing every boundary would cut a hairline segment off the end that reads as a rendering fault.
+ * 0.5% of a 20 second file is 100ms, of a two hour file 36 seconds.
+ */
+const EDGE_FRACTION = 0.005
+
+const OPAQUE = '#000'
+const CLEAR = '#0000'
+
+/**
+ * Where the bar is broken, as percentages, including the two ends: [0, ...breaks, 100].
+ *
+ * Both edges of every chapter count, so chapters that leave un-named time between them break the
+ * bar on each side of the gap and the un-named span becomes a segment of its own. Returns nothing
+ * when there is no break worth drawing, which is what leaves a file with no chapters, or one whose
+ * single chapter spans the whole picture, rendering exactly as it did before chapters existed.
+ */
+export const segmentBounds = (chapters: MediaChapter[], duration: number): number[] => {
+  if (!duration) return []
+  const breaks = new Set<number>()
+  for (const { start, end } of chapters) {
+    for (const seconds of [start, end]) {
+      const fraction = seconds / duration
+      if (fraction <= EDGE_FRACTION || fraction >= 1 - EDGE_FRACTION) continue
+      // rounded so two chapters meeting at the same instant cannot produce two boundaries a
+      // billionth apart, which would draw a double gap
+      breaks.add(Number((fraction * 100).toFixed(4)))
+    }
+  }
+  if (!breaks.size) return []
+  return [0, ...[...breaks].sort((a, b) => a - b), 100]
+}
+
+/**
+ * A mask painting the segments `keep` accepts, with a gap at every boundary between segments.
+ *
+ * The gap is cut from both sides of a boundary so it stays centred on it, and never from the two
+ * outer edges, where it would shorten the bar rather than divide it.
+ */
+export const segmentMask = (bounds: number[], keep: (index: number) => boolean): string => {
+  const half = CHAPTER_GAP_PX / 2
+  const last = bounds.length - 2
+  const stops: string[] = []
+  for (let i = 0; i <= last; i += 1) {
+    const from = bounds[i]!
+    const to = bounds[i + 1]!
+    const left = i === 0 ? '0%' : `calc(${from}% + ${half}px)`
+    const right = i === last ? '100%' : `calc(${to}% - ${half}px)`
+    stops.push(`${keep(i) ? OPAQUE : CLEAR} ${left} ${right}`)
+    if (i !== last) stops.push(`${CLEAR} calc(${to}% - ${half}px) calc(${to}% + ${half}px)`)
+  }
+  return `linear-gradient(90deg, ${stops.join(', ')})`
+}
+
+/**
+ * What a chapter is, when its title says plainly enough to offer a skip.
+ *
+ * Only ever a suggestion: the player shows a button for a few seconds and does nothing unless it is
+ * pressed. That is what lets this be generous rather than careful. A chapter wrongly called an
+ * opening costs a button nobody presses, while an opening this fails to recognise costs the feature.
+ *
+ * Every rule below was checked against 192 real files (891 chapter markers, 79 distinct chapter
+ * sequences) rather than guessed, and the counts quoted are from that sample.
+ */
+export type ChapterKind = 'opening' | 'ending'
+
+/**
+ * Labels that name the thing outright.
+ *
+ * "Credits" is here rather than among the hedged words because the sample settles it: 48 markers,
+ * a median length of 90.0s, and the last chapter of its file every time.
+ */
+const OPENING_LABELS = new Set([
+  'op', 'ops', 'opening', 'openings', 'ncop',
+  'オープニング', 'op主題歌',
+])
+const ENDING_LABELS = new Set([
+  'ed', 'eds', 'ending', 'endings', 'nced', 'credits', 'endcard', 'endroll',
+  'エンディング', 'エンドカード', 'エンドロール', 'ed主題歌',
+])
+
+/**
+ * Used for BOTH the theme and the scene before it, so accepted only when the file has not already
+ * named a plain one of that kind.
+ *
+ * "Intro" is the case that matters: 64 markers in the sample. Where the file also carries an OP or
+ * Opening it runs 62.9s to 240s and is a cold open; where it does not, it runs 3s to 91s and is the
+ * theme. Deferring to the plain marker separates the two without having to guess from length.
+ */
+const HEDGED_OPENING = new Set(['intro', 'イントロ'])
+/**
+ * "Outro" hedges for the same reason, found the same way: the one file in the sample that uses it
+ * reads Episode, Credits (90.1s), Outro (56s), Preview. The 90s Credits is the theme and the Outro
+ * is the scene after it, so a plain marker in the same file has to win.
+ *
+ * "Epilogue" is deliberately absent. It is never the theme in the sample (4s in eight files, then
+ * 16s, 56s and 122s), so it is a post-ending scene rather than a hedged name for one.
+ */
+const HEDGED_ENDING = new Set(['outro', 'アウトロ'])
+
+/**
+ * Whole labels decided outright, before any word is reduced away.
+ *
+ * "Credits Start" and "Credits End" are one release's way of bracketing the ending, and they are not
+ * interchangeable: the first IS the theme (19 markers, median 89.0s), the second is the 28s tail
+ * after it. Reduced word by word the two would come out identical, so they are read whole.
+ */
+const EXACT: Record<string, ChapterKind | null> = {
+  'credits start': 'ending',
+  'credits end': null,
+  'preview end': null,
+  'end credits': 'ending',
+  'closing credits': 'ending',
+}
+
+/** Carried alongside a label without changing what it names. */
+const MODIFIERS = new Set([
+  'theme', 'themes', 'song', 'sequence',
+  'nc', 'non', 'credit', 'creditless', 'textless', 'clean',
+  'tv', 'size', 'version', 'ver', 'full', 'the',
+  '主題歌', 'ノンクレジット',
+])
+
+/**
+ * Words that turn a label into something that comes AFTER the thing, not the thing.
+ *
+ * "Post-Credits" is a real chapter in the sample, sitting between the ED and the preview. Reading
+ * only its last segment it is indistinguishable from "Credits", which is why the segments before a
+ * match get a look too.
+ */
+const NEGATORS = new Set(['post', 'pre', 'after', 'before', 'non'])
+
+/** Everything before the first separator: releases write "OP - Song Name" and "Opening: Title". */
+const LABEL_SEPARATORS = /[-:|~/\\[\]()「」『』【】,.!?"'’]|\s+by\s+/u
+
+/** Titles a container writes when it has nothing to say, which must never look like a marker. */
+const isPlaceholder = (title: string): boolean =>
+  // "Chapter 07", and the timestamps one muxer writes as titles, which are 300 markers in the sample
+  /^chapter\s*\d+$/i.test(title.trim()) || /^\d{1,2}:\d{2}:\d{2}([.,]\d+)?$/.test(title.trim())
+
+const normalise = (title: string): string =>
+  (title.trim().toLowerCase().split(LABEL_SEPARATORS)[0] ?? '').trim().replace(/\s+/gu, ' ')
+
+/**
+ * The label reduced to the words that name it, or nothing when it does not reduce to one.
+ *
+ * Numbers go, so OP2 and "Opening 2" are both an opening, and decoration goes, so is "Opening Theme".
+ * What has to be left is a SINGLE word, and that requirement is the whole guard against a title that
+ * merely mentions a theme: "The Ending of Everything" keeps four words and matches nothing.
+ */
+const labelWords = (label: string): string[] =>
+  label
+    // a trailing number belongs to the label rather than being a word: op1, ed02, opening 3
+    .replace(/([a-z぀-ヿ一-鿿])\s*\d+\s*$/u, '$1')
+    .split(/[\s_]+/u)
+    .map((word) => word.replace(/^[^\w぀-ヿ一-鿿]+|[^\w぀-ヿ一-鿿]+$/gu, ''))
+    .filter((word) => word && !/^\d+$/.test(word) && !MODIFIERS.has(word))
+
+const ofWord = (word: string, hedged: boolean): ChapterKind | undefined => {
+  if (OPENING_LABELS.has(word)) return 'opening'
+  if (ENDING_LABELS.has(word)) return 'ending'
+  if (!hedged) return undefined
+  if (HEDGED_OPENING.has(word)) return 'opening'
+  if (HEDGED_ENDING.has(word)) return 'ending'
+  return undefined
+}
+
+const kindOf = (title: string, hedged: boolean): ChapterKind | undefined => {
+  if (isPlaceholder(title)) return undefined
+  const whole = title.trim().toLowerCase().replace(/\s+/gu, ' ')
+  if (whole in EXACT) return EXACT[whole] ?? undefined
+
+  /*
+   * The label is usually first, but a release will also hang it off the end: "Song Name (Opening)"
+   * and "Song Name - OP" are both real. Both ends are tried, and each has to reduce to one naming
+   * word on its own.
+   *
+   * A SEPARATOR is what makes the second one safe. Without it, reading the last word of a title
+   * would skip a real scene called "Proclamation of a meeting opening", which is an actual chapter
+   * in a real release. That title has no separator, so it stays one many-word segment and matches
+   * nothing.
+   */
+  const segments = whole.split(LABEL_SEPARATORS).map((part) => part.trim()).filter(Boolean)
+  if (segments.some((segment) => segment.split(/[\s_]+/u).some((word) => NEGATORS.has(word)))) return undefined
+
+  const ends = segments.length > 1 ? [segments[0]!, segments.at(-1)!] : segments
+  for (const segment of ends) {
+    const words = labelWords(segment)
+    if (words.length !== 1) continue
+    const kind = ofWord(words[0]!, hedged)
+    if (kind) return kind
+  }
+  return undefined
+}
+
+/**
+ * Below this there is nothing worth offering to skip, in seconds.
+ *
+ * Some releases write a marker rather than a span: one group labels a 3s beat "Intro" and a 2s tail
+ * "Credits", and a button that jumps you forward three seconds reads as broken. The sample leaves a
+ * clean gap to cut in: every junk marker in it is 5s or shorter, and the shortest real theme is 26s
+ * (a shortened credits sequence). Anything in between would do; 15 is the middle of the gap.
+ */
+const SHORTEST_WORTH_SKIPPING = 15
+
+/**
+ * How long a theme runs, in seconds, for a file whose chapters carry no usable name.
+ *
+ * Ninety seconds, give or take three, because the standard really is that tight: the sample gives a
+ * median of 89.9s for Opening, 90.0 for OP, 90.0 for ED and 89.5 for Ending.
+ *
+ * Chosen by sweeping both populations, which want opposite things. Widening it looks better against
+ * the files that HAVE titles, where the answer is known: 70 to 110 agreed on 137 segments against
+ * 123 here. But those files never reach this rule. On the 53 files that do, a wide band pulls in a
+ * neighbouring chapter that happens to be 94s, leaves two candidates in the same half, and abstains:
+ * 70 to 110 finds an opening in 44 of them where this finds one in 49.
+ *
+ * Every band quoted above contradicts a title nowhere. Being narrow is what keeps it that way.
+ */
+const THEME_MIN = 87
+const THEME_MAX = 93
+
+/**
+ * A guess from shape alone, for a file whose chapters are named by a muxer rather than by a person.
+ *
+ * 53 files in the sample name every chapter after its own timestamp or call it "Chapter 07", which
+ * says nothing at all. 46 of those carry exactly one theme-length chapter in each half of the
+ * runtime, which is the whole rule: one candidate in the first half is the opening, one in the
+ * second is the ending, and anything less clear than that is left alone.
+ *
+ * Ambiguity is resolved by abstaining, never by picking. Two candidates in a half means neither is
+ * offered, because guessing between them is how a viewer gets thrown out of the episode.
+ */
+const guessFromShape = (chapters: MediaChapter[]): (ChapterKind | undefined)[] => {
+  const blank: (ChapterKind | undefined)[] = chapters.map(() => undefined)
+  // the last chapter's end stands in for the runtime, which is all these files ever have
+  const runtime = chapters.at(-1)?.end ?? 0
+  if (!runtime) return blank
+
+  const candidates = chapters
+    .map((chapter, index) => ({ index, start: chapter.start, length: chapter.end - chapter.start }))
+    .filter(({ length }) => length >= THEME_MIN && length <= THEME_MAX)
+
+  const half = runtime / 2
+  const first = candidates.filter(({ start }) => start < half)
+  const second = candidates.filter(({ start }) => start >= half)
+  if (first.length === 1) blank[first[0]!.index] = 'opening'
+  if (second.length === 1) blank[second[0]!.index] = 'ending'
+  return blank
+}
+
+/**
+ * What each chapter is, decided across the whole list rather than one title at a time.
+ *
+ * The list is what resolves the hedged words. A file whose chapters read Intro, OP, Episode, ED,
+ * Preview has already said which one the theme is, so its Intro is left alone; one that reads
+ * Episode, Intro, Episode, Credits has not, so its Intro is offered. Both shapes are in the sample,
+ * 13 files and 24 files.
+ */
+export const classifyChapters = (chapters: MediaChapter[]): (ChapterKind | undefined)[] => {
+  const plain = chapters.map(({ title }) => kindOf(title, false))
+
+  /*
+   * A file that is MOSTLY themes is a creditless bonus disc, where they are the content.
+   *
+   * Offering to skip an opening on a disc of nothing but openings is the one failure this design
+   * cannot shrug off, because the button would invite the viewer past the exact thing they put on.
+   *
+   * Measured by RUNTIME rather than by how many chapters match. Counting them suppresses an
+   * ordinary episode, which is two themes out of three chapters; by runtime that same episode is
+   * 13% theme and a creditless disc is all of it.
+   */
+  const themed = chapters.reduce((sum, c, i) => sum + (plain[i] ? c.end - c.start : 0), 0)
+  const total = chapters.reduce((sum, c) => sum + (c.end - c.start), 0)
+  if (total > 0 && themed * 2 > total) return chapters.map(() => undefined)
+
+  const named = chapters.map((chapter, i) => {
+    if (chapter.end - chapter.start < SHORTEST_WORTH_SKIPPING) return undefined
+    const certain = plain[i]
+    if (certain) return certain
+    const guess = kindOf(chapter.title, true)
+    return guess && !plain.includes(guess) ? guess : undefined
+  })
+
+  // shape is the last resort, and only for a file whose titles said nothing whatsoever
+  return named.some(Boolean) ? named : guessFromShape(chapters)
+}

package/src/lib/utils/source.ts

@@ -24,8 +24,19 @@ const fromUrl = async (
 ) => {
   const length = _length ?? await probeLength(url, credentials)
   const read = async (offset: number, size: number) => {
+    /*
+     * A read that starts at or past the end answers empty, the way slicing a Blob does.
+     *
+     * Without this the range comes out as `bytes=<offset>-<length-1>` with the last byte before the
+     * first, which is not a range a server can satisfy: it answers 416 and the read throws, so the
+     * demuxer's ordinary walk into EOF surfaces as "Reading the video file failed" on a file that is
+     * perfectly fine. The blob arm has always returned empty here, so this is also what makes the two
+     * arms of `inputToRemuxerInput` behave the same.
+     */
+    const end = Math.min(offset + size, length) - 1
+    if (end < offset) return new ArrayBuffer(0)
     const response = await fetch(url, {
-      headers: { Range: `bytes=${offset}-${Math.min(offset + size, length) - 1}` },
+      headers: { Range: `bytes=${offset}-${end}` },
       credentials,
     })
     if (!response.ok) throw new Error(`The source could not be read: HTTP ${response.status}`)