@autoaiai/aicut 0.1.11 → 0.1.13

package/skills/html2video/references/narration.md

@@ -1,111 +0,0 @@
-# 口播模式：写给人念的稿子
-
-口播模式不是"音乐版再附一个文档"。它是**另一种剪法**，因为承载内容的东西换了：
-音乐版里屏幕文字就是内容，口播版里声音才是内容，文字退回去当锚点。
-
-三件事随之改变：
-
-| | 音乐版 | 口播版 |
-|---|---|---|
-| 时长由什么决定 | 默读速度（可读性下限） | **语速**（旁白长度） |
-| 屏幕文字 | 完整承载信息 | **只留标题和数字**，图注多半要删 |
-| `target.seconds` | 求解目标 | **只是断言**，记录实际长度 |
-
-启用方式：`audio.mode: "voice"`，并给**每一个**场景写 `narration`。schema 会强制后者。
-然后：
-
-```bash
-node scripts/render.mjs --storyboard ./work/myrun --script    # 只出稿
-node scripts/render.mjs --storyboard ./work/myrun             # 出稿 + 渲染
-```
-
-稿子落在 `narration.md`。voice 模式下渲染会自动出稿，因为不给稿子的口播成片是没法录的。
-
-## 最重要的一条：不要把屏幕上的字念出来
-
-这是口播视频最常见、也最致命的毛病。观众正在读标题，你同时把同一句话念出来，
-两个通道传的是同一份信息 —— 结果是两边都记不住，而且显得敷衍。
-
-屏幕负责**说是什么**，旁白负责**说为什么、所以呢、哪来的**。
-
-| 画面标题 | ❌ 复述 | ✅ 补充画面没说的 |
-|---|---|---|
-| 把思维翻译成人话 | "这个方法能把思维翻译成人话。" | "Anthropic 最近公开了一个方法，冲的是可解释性里一个很基本的麻烦。" |
-| 一个负责说，一个负责猜 | "一个模型负责说，一个负责猜。" | "方法本身不复杂，难的是怎么知道翻译得对不对。他们的办法是训两个模型……" |
-| NLA 也会编造 | "NLA 也是会编造的。" | "但这套方法有个必须说清楚的毛病。" |
-
-生成的稿子里每一场都有「画面」一栏，写着那一刻屏幕上有什么。它在那里是**为了让你避开它**。
-
-## 屏幕文字要收
-
-旁白讲得完的东西，不必再写在画面上。实践下来：
-
-- `figure` 场景的 `caption` 通常整条删掉 —— 那句话交给旁白说
-- `sub` 保留一到两行，别三行
-- `stat` 的大数字和 `of` 留着（数字看比听准）
-- `ladder` 的条目留着，但旁白要把它们**串起来**讲，不是逐条念
-
-校验器会替你抓这个：如果某一场屏幕文字的默读时间超过旁白时长，它会报
-
-```
-scene "s7": its on-screen text needs 3.0s longer to read than the narration lasts.
-```
-
-意思是这一场你在要求观众同时读和听。两个解法：把文字砍掉，或者把旁白写长
-（如果画面上的内容确实值得那么多时间）。两条都对，取决于那些字该不该留。
-
-## 口语句法，不是书面语
-
-念出来和读进去是两种语言。稿子要按念的写：
-
-- **短句**。一个逗号一个呼吸。书面语那种层层嵌套的定语，念出来会断气。
-- **不要"其"、"之"、"该"、"上述"**。念着别扭。
-- **数字口语化**。写 "26%" 屏幕上没问题，念的时候是"四分之一左右"更顺 ——
-  除非精确值本身是重点。
-- **专有名词保留原文**。NLA、SWE-bench、Opus 4.6 照写，念的人自然会处理。
-- **用 `\n` 断句**。稿子会把它保留成换气点，也是画面重排的依据。
-
-沿用 [editorial.md](editorial.md) 的禁用词表：深入探讨、让我们看看、带你了解、
-一探究竟、惊艳、革命性、总而言之。念出来比写出来更假。
-
-## 时长是怎么算的
-
-没有音频可量，所以从文字估。估的是**音节**不是字符，这个区别很关键 ——
-中文语速按音节稳定，而字符到音节的映射对这个管线偏爱的素材恰好不是 1:1：
-
-```
-"26%"        3 字符  →  百分之二十六        6 音节
-"1 Tbps"     6 字符  →  一 T-b-p-s ≈        4-5 音节
-"SWE-bench"  9 字符  →  ≈                   5 音节
-```
-
-按字符数估会把数据密集的稿子**严重估短**，每个数字镜头都会被切掉尾巴。所以
-`lib/speech.ts` 做了展开：数字串 n 位算 `n + n/2` 音节，`%` 算 3，拉丁串算长度一半。
-
-默认 **4.3 音节/秒**（约 258 字/分钟），是常规解说的语速。误差约 ±15%，由每场的
-尾留吸收。你的语速明显不同就换：
-
-```bash
-node scripts/render.mjs --storyboard ./work/myrun --script --sps 3.8   # 慢一点
-```
-
-画面会跟着重排，稿子里的语速假设也会更新。
-
-## 稿子里有什么
-
-`narration.md` 分三部分：
-
-1. **头部** —— 成片时长、总字数、语速假设、录制须知
-2. **逐场** —— 时间码（互不重叠的连续窗口，可分段重录）、那一刻的画面、
-   时长预算（旁白 + 余量）、要念的那句话
-3. **连读版** —— 去掉一切标注的纯旁白，给提词器用
-
-余量低于 0.2s 的场景会单独标出来，那意味着念的时候不能停顿 —— 通常该砍几个字。
-
-## 已知边界
-
-- **稿子不含音频。** 你自己录，或者外接 TTS 后用 `voiceDurations` 把实测时长喂回
-  `resolveTimeline`（函数签名已经留好，接线还没做）。
-- **没有烧字幕。** 字幕需要词级时间戳，那要么来自 TTS，要么要跑强制对齐 —— 属 P1。
-- **时长是估的。** 实际录音会和排布差几秒。每场的尾留是为此留的缓冲，但如果你
-  的语速和 4.3 音节/秒差得远，先用 `--sps` 校准一次再录。

package/skills/html2video/references/pitfalls.md

@@ -1,238 +0,0 @@
-# Pitfalls
-
-Every entry here cost a debugging cycle or a bad render. Read this before editing
-`assets/template/src/`, and add to it when something new bites.
-
-## Contents
-
-- [Remotion's rendering model](#remotions-rendering-model)
-- [Things the official docs get wrong](#things-the-official-docs-get-wrong)
-- [Props and the silent wrong-video failure](#props-and-the-silent-wrong-video-failure)
-- [Rendering reliability](#rendering-reliability)
-- [Assets and image quality](#assets-and-image-quality)
-- [Audio frame spaces](#audio-frame-spaces)
-- [Layout failures](#layout-failures)
-- [Determinism](#determinism)
-
----
-
-## Remotion's rendering model
-
-Frames are rendered by several browser tabs in parallel, and **those tabs share no
-state**. Everything else follows from that.
-
-**CSS animations, transitions and Tailwind `animate-*` are forbidden.** They do
-not render, or worse, they render differently per tab and the video flickers
-mid-shot. Animate only from `useCurrentFrame()`.
-
-Harvested pages are full of `@keyframes`, `transition:` and `animate-*`, so
-harvested CSS is **never passed through as CSS**. `harvest.py` reduces the page's
-styling to a flat JSON of primitives via `getComputedStyle`, and `zArt` validates
-that against an allow-list with type and range checks. There is no path by which
-page CSS reaches the render.
-
-**`useState`/`useEffect` for anything visual is a bug**, for the same reason: the
-effect timing differs per tab. The one exception is `WaitForFonts`, which uses
-them to gate the whole composition, and which is allowed by name.
-
-## Things the official docs get wrong
-
-The bundled `remotion-best-practices` skill is good on timing, sequencing,
-transitions and audio, and is worth reading. But it has errors and gaps that
-matter here.
-
-**`interpolate()` cannot return a string.** Verified against
-`node_modules/remotion/dist/cjs/interpolate.d.ts`:
-
-```ts
-export declare function interpolate(input: number, inputRange: readonly number[],
-  outputRange: readonly number[], options?: InterpolateOptions): number;
-```
-
-The skill's own example — `translate: interpolate(frame, [0,100], ["0px 0px", "100px 100px"])`
-— does not typecheck. Wrap a numeric interpolate in a template literal instead.
-
-**`translate` needs the two-value form.** `translate: 26` is translateX only, so a
-"rise" animation written that way slides sideways. Always `translate: \`0 ${y}px\``.
-`src/lib/shots/primitives.tsx` is the reference for the correct pattern.
-
-**Ken Burns is not covered at all** — no pan, no zoom, no overscan. All of it is
-in `src/lib/move.ts` and `src/lib/shots/KenBurnsImage.tsx`.
-
-**`iris` and `none` transitions exist** in `@remotion/transitions` 4.0.409 even
-though the skill documents neither. `clockWipe` and `iris` require explicit
-`width`/`height` props; `fade` has a `shouldFadeOutExitingScene` flag that turns a
-fade-over into a real cross-dissolve (this skill sets it).
-
-**`delayRender()`'s 30-second default timeout is never mentioned.** This project
-allows exactly three `delayRender` sites — the font handle, `<Img>` decode, and
-`<Audio>` in voice mode — and forbids `http://` or `https://` anywhere in `src/`
-or in any `asset.src`, so nothing in the render waits on the network. That
-invariant is why the timeouts are generous rather than tight.
-
-**CJK font guidance is absent and the examples are actively wrong** for Chinese:
-every `loadFont` example hardcodes `subsets: ["latin"]`, which renders Chinese as
-tofu. See [cjk-type.md](cjk-type.md).
-
-## Props and the silent wrong-video failure
-
-**Remotion MERGES `--props` with `defaultProps`.** The composition's props are
-shaped `{ storyboard: … }`. Passing `storyboard.json` directly therefore leaves
-`props.storyboard` pointing at the built-in demo board, and Remotion renders the
-DEMO with your board's keys sitting unused alongside it.
-
-This fails silently and produces a plausible video. It cost two rounds of
-"why didn't my change take effect".
-
-`scripts/render.mjs` wraps props automatically. Use it. If you must call Remotion
-directly, the props file has to look like `{"storyboard": { … }}`.
-
-Related: regenerate the props file after every storyboard edit. `render.mjs` does
-this each run, which is the actual fix — a stale intermediate is not a mistake
-worth remembering, it is one worth deleting.
-
-## Rendering reliability
-
-**Concurrency above ~4 fails against the system Chrome.** Symptom:
-
-```
-Error: Visited "http://localhost:3001/index.html" but got no response.
-  at getPool (…/render-frames.js)
-```
-
-One of N parallel tabs never loads. Measured here: 8 fails, 3 is stable. Stills
-always work because they use one tab. `render.mjs` defaults to 8 — pass
-`--concurrency 3` when it bites.
-
-Remotion prefers its own `chrome-headless-shell`, which is more reliable for
-rendering than a full Chrome. Note that `npx remotion browser ensure` will NOT
-download it if `remotion.config.ts` already sets `setBrowserExecutable()` — it
-reports "Has browser at …" and exits. To force the download, comment out the
-browser-executable branch first.
-
-**Composition duration comes from `calculateMetadata`.** Asking for a frame beyond
-it fails with `RangeError: Cannot use frame N: Duration of composition is M`. The
-validation report prints each scene's `from` frame; use it to pick still frames.
-
-## Assets and image quality
-
-**Image proxies silently degrade figures.** Next.js sites serve
-`/_next/image?url=<original>&w=3840&q=75`. A fetcher that follows the rendered
-`<img src>` gets the proxied file. Measured on the reference article: a
-4620x1410 source came back as **3840x1172 and palette-quantised to 256 colours**.
-
-`harvest.py` decodes the `url=` parameter and downloads the original. Since this
-pipeline's premise is that figures have spare pixels to push into, that 20% and
-the colour depth both matter. If you add a fetch path, preserve this behaviour.
-
-**Never upscale.** `KenBurnsImage` fixes the layout at the tightest zoom so
-`scale` never exceeds 1 and Chrome only ever downsamples. The schema additionally
-rejects any crop needing more source pixels than the image has, before frame 0.
-A soft hero shot looks fine in a thumbnail and obviously wrong at 1080p.
-
-**Zoom range is capped by duration** (`maxZoomRatio`): 1.25x for moves of 4s or
-more, 1.6x for short ones. Chrome re-rasterises across large scale ranges, and on
-a slow move that shimmers.
-
-**Intermediate frames are PNG, not JPEG.** `jpeg` is the right default for video
-footage and is what other Remotion projects set, but this pipeline renders flat
-colour fields and fine CJK strokes — precisely where JPEG's DCT ringing appears
-along glyph edges and where chroma subsampling eats a 2px accent rule.
-
-Measured, rather than assumed. Rendering scene s4 both ways and comparing against
-a `remotion still` of the same frame as ground truth:
-
-| intermediate | SSIM | segment size |
-|---|---|---|
-| jpeg | 0.9874 | 149.7 kB |
-| png | **0.9944** | 159 kB |
-
-Error roughly halved for 6% more bitrate. The cost of PNG is render time and temp
-disk, not quality, so it is the default here. Revisit only if a board is dominated
-by photographic material.
-
-## Audio frame spaces
-
-**`<Audio volume={(f) => …}>` receives AUDIO-LOCAL frames, not composition
-frames.** Remotion derives `f` from `useMediaStartsAt()`, so if the `<Audio>` is
-inside a `<Sequence from={90}>` or carries a `trimBefore`, every duck window
-computed in composition space is wrong by that offset — and nothing errors.
-
-The rule that removes the problem rather than managing it: **mount the bed at
-composition frame 0, outside any Sequence, with no `trimBefore`.** Then the two
-frame spaces coincide. If an offset is ever genuinely needed, pass it explicitly
-and add it inside `duckAt` — never infer it from the callback.
-
-## Layout failures
-
-**Do not mix absolutely positioned images with flex text.** The first version of
-`FigureReveal` did, and headlines landed on top of figures. Once the image is a
-flow child in the same layout pass, overlap becomes impossible to write rather
-than something to remember. Full-bleed is the deliberate exception: there the
-image IS the background and text sits over a scrim.
-
-**Do not crop into a wide schematic.** The first `DiagramWalk` panned across a
-3.28:1 diagram at 55% width. That destroyed the left-to-right flow that was the
-diagram's entire content, and magnified the figure's own labels until they
-outweighed the headline. A schematic wants to be shown whole, at the largest size
-that fits, with attention moved by annotation. `move: hold` over the full rect is
-the intended default.
-
-**Do not put callout chips on top of a diagram.** A diagram is already dense with
-type. Chips belong in the margin the layout reserves for them, with a short leader
-up to the region they name.
-
-**Rules thinner than 2px do not exist in video.** A harvested `ruleWidth: 1` is a
-CSS hairline designed for a browser at arm's length; after H.264 and arbitrary
-viewing scale it vanishes. Measured: a 1px title rule was present in the pixels
-and invisible to the eye. `palette()` scales harvested widths up to
-`hairline` (min 2px) and `ruleAccent` (min 4px) once, centrally.
-
-**A dark scrim over a light figure on a light ground reads as a mistake.** Scrim
-direction follows the ground's luminance (`isDark()`), and scrims are only used
-for full-bleed.
-
-**Do not keep a second table of box dimensions.** `SLOT_PX` in `lib/design.ts` is
-the only one. An earlier version added a parallel `FLOW_BOX` for render geometry
-with different heights, so the resolution guard aspect-normalised the crop to
-1.78:1 while the renderer normalised it to 2.38:1 — the guard validated a rectangle
-that was never displayed, and its "needs Npx" figure described pixels nobody
-sampled. Nothing failed; the numbers were just quietly meaningless.
-
-**Every slot is 1.78:1 or wider, so a taller figure loses height under `cover`.**
-On a chart that height is the title and the axis labels. This was found by a cold
-run on a chart-heavy article, where the operator resorted to pre-padding PNGs
-externally to force the aspect. Use `"fit": "contain"` instead — see
-[storyboard.md](storyboard.md#fit-cover-or-contain).
-
-**Centre the image+text group in a non-full-bleed figure, not its contents.** Both
-children are the slot's width, so left-anchoring the column leaves all the slack on
-one side and a 1240px chart in a 1920px frame looks accidentally off-centre.
-
-**The character-budget check must be per shot.** Assuming "headline at 96px on two
-lines" everywhere was wrong in both directions: it rejected legal copy on
-`statement` and `outro` (which set the headline at 72px on three lines) and waved
-through text far too long for a `figure` in a `left`/`right` slot, whose real box is
-800px. `TEXT_BOX` in `lib/design.ts` holds the per-shot boxes; keep it in sync with
-what the shots pass to `useFits`.
-
-## Determinism
-
-Two symptoms, one root cause. Parallel-tab disagreement shows up as mid-shot
-flicker; run-to-run disagreement shows up as an unreproducible render.
-
-- `remotion`'s `random(seed)` is the only permitted randomness.
-- `src/lib/sandbox.ts` — imported first from `Root.tsx` — hard-overrides
-  `Math.random`, `Date.now` and `performance.now` while rendering. Overrides
-  rather than lint rules, because a lint rule cannot see a transitive dependency.
-- `Intl.Segmenter` is deliberately **off** for layout. It works, and it would give
-  slightly nicer Chinese line breaks, but ICU word-break data varies between
-  Chrome builds — so the same storyboard could break lines differently on two
-  machines. Per-character plus kinsoku is self-contained and provably identical.
-- Fonts are subset locally from pinned sources, so metrics are byte-identical
-  everywhere. If the download fails and the skill falls back to a macOS system
-  font, it says so — and that guarantee is gone.
-
-Anything unloaded when `measureText` runs silently returns Latin-fallback metrics,
-which bakes in the wrong line breaks. `fitCJK` passes
-`validateFontIsLoaded: true` so that throws instead of lying.

package/skills/html2video/references/shots.md

@@ -1,198 +0,0 @@
-# The shot library
-
-Ten shots. Each owns its layout, its entrance choreography and its emphasis
-handling; the storyboard supplies content and intent only.
-
-Two rules run through all of them, and they are what keeps the output from looking
-like slides:
-
-- **Readable content sits in flow layout inside the safe area.** Absolute
-  positioning is for backgrounds, scrims and decoration. Reserve a slot for every
-  element and animate it from that slot — never into space another element occupies.
-- **No cards, badges, pills or tiny labels.** Those are web UI patterns; in video
-  they read as clutter.
-
-## Entrance primitives
-
-Everything enters through one of two primitives, so timing stays consistent.
-
-**`MaskUp`** — text rises from behind a hard edge, clipped to exactly one line
-height. Reads as typeset rather than animated, which is why it is the default for
-headlines. Selected by `motion.revealStyle: "mask-up"` or `"clause-cascade"`.
-
-**`Rise`** — fade plus a short upward translate, with optional blur. The default
-for anything that is not a line of type.
-
-**`Rule`** — a hairline that draws itself via `scaleX` from a left or right origin.
-Used for sweeps, underlines, strikes and row dividers.
-
-Lines within a block stagger by `PACE[pace].stagger` (4 / 7 / 11 frames for
-staccato / measured / languid). Entrance duration is `PACE[pace].enter`.
-
-## Emphasis
-
-Authored as `{ at: 0..1, kind, target }` and resolved to frames against the solved
-scene duration, so re-timing can never desync it.
-
-| kind | Effect |
-|---|---|
-| `punch` | brief `scale` to 1.035 and a tint to accent, returning to 1 — reads as stress, not as a layout change |
-| `flash` | the ground dips to 0.82 opacity for 6 frames; the frame flinches |
-| `underline` | an accent rule sweeps beneath the target |
-| `strike` | a rule draws across the target, tilted 0.6 degrees so it reads as struck by hand rather than as a border |
-| `rule-sweep` | times the shot's structural rule to a beat instead of to the entrance |
-| `callout` | an accent rule plus dot marking a focal point |
-| `desaturate` | drops the target to muted colour |
-
-`target` is `"headline"`, `"sub"`, or `"line:N"` (1-based) to hit a single line.
-
-## The shots
-
-### `title`
-Opening. Sets register, subject and source in one frame.
-
-Eyebrow top-left in accent; display headline on `cols(8)`, left-aligned; an accent
-rule on `cols(4)`; Latin sub beneath in muted. Headline reveals per line by
-`MaskUp`; the rule draws on the `rule-sweep` beat if one is authored, otherwise
-after the headline lands.
-
-### `statement`
-A single claim with nothing to compete against it. Centred column, `cols(10)`, at
-most three lines at statement size. Nothing else in the frame — the emptiness is
-the design. Use `punch` on the line that carries the turn.
-
-### `stat`
-One number made the subject. Value at 240px tabular in the display face, unit at
-52px in accent, baseline-aligned; `of` caption beneath in `cols(9)`.
-
-Digits cascade in 3 frames apart. Deliberately character-by-character rather than a
-count-up, because a count-up breaks on non-numeric values like "3x" or "<1%".
-
-### `quote`
-A sentence worth quoting verbatim. The opening 「 hangs outside the text box
-(`marginLeft: -1em`) — see [cjk-type.md](cjk-type.md) for why this matters.
-Quote on `cols(9)`, attribution right-aligned beneath.
-
-### `figure`
-One self-contained finding and its image. Behaviour depends on the slot:
-
-- **`full`** — the image IS the background. Ken Burns over the whole frame, text at
-  the bottom over a scrim whose direction follows the ground's luminance.
-- **`inset` / `band` / `left` / `right`** — the image is a flow child in a reserved
-  box with a hairline edge, and the text sits in its own row beneath. No scrim, ink
-  on paper.
-
-Use `push-in` to settle onto a detail, `pull-out` to reveal context.
-
-### `compare`
-Two states of one thing. Headline, optional figure, then two labelled columns of
-`cols(6)` with an accent divider that draws downward between them.
-
-The `underline` emphasis sweeps across **both** columns rather than under each,
-because the point of the shot is the relationship, not the pair.
-
-### `diagram`
-A mechanism, walked through one callout at a time.
-
-**Default to `slot: "band"`, `"fit": "contain"`, and `move: hold` over the full
-rect.** All three parts matter. Cropping a wide schematic destroys the
-left-to-right flow that IS the diagram's content and magnifies the figure's own
-labels until they outweigh the headline — so show it whole and move attention by
-annotation instead.
-
-`"fit": "contain"` is the part that is easy to leave out, and leaving it out
-quietly re-introduces the crop: `band` is 3.28:1, so under the default `cover` a
-2.5:1 diagram still loses height off the top and bottom — usually its subtitle and
-its legend. `hold` over the full rect does not prevent that; only `contain` does.
-
-Callout chips sit **below** the band, positioned at their anchor's projected x,
-with a short accent leader up to the region they name. Exactly one is at full
-strength; spent ones drop to 42% so the thread of what was already said stays
-readable.
-
-Panning remains available for figures genuinely too large to read at once: give
-callouts a `focus` rect and the camera retargets between them, landing a beat
-before the words arrive so the eye is already there.
-
-### `caveat`
-The limitation. **The one inverted shot in the system** — ink ground, paper text.
-That is deliberate: the tonal break makes the caveat land structurally before it is
-read, and it is why this shot needs no fade in (cut into it).
-
-Headline off-centre on `cols(7)`, sub beneath with a tilted hairline available for
-`strike`. Use `flash` early and `strike` late.
-
-### `ladder`
-Three or four takeaways. Numbered rows: label in accent at 30px, text at sub size,
-hairline divider above each row.
-
-The rule draws four frames **before** its text arrives — the line appears, then the
-content lands on it. That small lead is what makes the sequence feel authored.
-
-### `outro`
-Core message centred at statement size; source label and URL small at bottom-left.
-A very slow `scale` to 1.012 across the whole shot — 1.2% is deliberately near the
-threshold of notice, so it reads as the film settling rather than as a move.
-
-## Asset slots
-
-Destination boxes in output pixels. These feed the resolution guard.
-
-| slot | box | aspect | Use for |
-|---|---|---|---|
-| `full` | 1920x1080 | 1.78:1 | photographs, textures, a figure that can carry a frame |
-| `band` | 1640x500 | 3.28:1 | wide schematics, timelines, anything 3:1 or wider |
-| `inset` | 1240x520 | 2.38:1 | screenshots, transcripts, charts — the ground shows around it |
-| `left` / `right` | 800x450 | 1.78:1 | side-by-side comparisons |
-
-These are the *only* slot dimensions; `SLOT_PX` in `lib/design.ts` is the single
-table, used by both the resolution guard and the renderer. (An earlier version kept
-a second table of render boxes with different heights, so the guard validated a
-rectangle that was never displayed. Do not reintroduce one.)
-
-**Every slot is 1.78:1 or wider, which matters a lot for charts.** Under the default
-`"fit": "cover"` an image taller than its slot loses the excess height — and on a
-chart that height is the title and the axis labels. A 1.36:1 map in `inset` loses
-~43% of its height. Set `"fit": "contain"` for anything whose edges carry meaning
-and the whole image is letterboxed instead, with the ground showing above and below.
-`contain` ignores `move`, since there is nothing to pan within.
-
-Pixel budget under `cover`: a slot needs `dest.w / tightest_crop.w` source pixels. A
-1590px-wide chart therefore **cannot** go in `full` at any crop (it would need
-1920+), and only fits `inset` at crops of 0.78 or wider. Under `contain` the demand
-is just the fitted size, so headroom is far easier to satisfy. The validator
-computes both cases and refuses what will not work.
-
-## Motion signature
-
-Seeded from the source URL, so two different articles move differently with no
-bespoke code. This — together with harvested colour and type, and the choice of
-shots — is where the perceived uniqueness comes from.
-
-```
-easeFamily   crisp | editorial | overshoot
-pace         staccato | measured | languid
-panBias      ltr | rtl | converge | settle
-revealStyle  mask-up | rise | scale-settle | clause-cascade
-gridEnergy   0..1   background drift amount
-transitionVocab  which transitions this video is allowed to use
-```
-
-Easing curves: `crisp` = `bezier(0.16, 1, 0.3, 1)` for decisive entrances,
-`editorial` = `bezier(0.45, 0, 0.55, 1)` for pans, `overshoot` =
-`bezier(0.34, 1.56, 0.64, 1)` for a slight bounce past target.
-
-Enter with ease-out, exit with ease-in.
-
-## Transitions
-
-Available: `cut`, `fade`, `wipe`, `slide`, `flip`, `clockWipe`, `iris`.
-
-`fade` is a true cross-dissolve (`shouldFadeOutExitingScene`), which reads as
-editorial; a fade-over reads as a slide deck. `cut` is the **absence** of a
-transition element, not a zero-frame one — total duration is
-`Σ scene frames − Σ transition frames`.
-
-Cut into tonal breaks (the `caveat`). Dissolve between related beats. Reserve
-`wipe`, `slide` and `clockWipe` for act boundaries; using them between every scene
-is what makes a video feel like a template.