Clone any viral video with AI agents. Not just a script, the whole workflow: swap the face, the words, the B-roll, ship 100 variants in one command, and get your 100M views.
TypeScript agentic-aiaiai-agentsai-videocompilerdslffmpeggenerative-aillmmarkup-languagemonorepoplugin-systemprogramming-languagetext-to-videotypescriptvideovideo-automationvideo-clonevideo-editingvideo-generation
docs: define flat track source architecture
kashorinfda65dc4f71a
diff --git a/README.md b/README.mdindex 12785617..52bd782c 100644--- a/README.md+++ b/README.md@@ -1,17 +1,17 @@# SVMLSemantic Video Markup Language is a semantic source language for information-flow-video. Script creates stable semantic addresses; a Composition-selected Temporal-Producer binds every Segment and word endpoint to one ProgramBasis; ordinary-imported Kernels project the resulting media graph to one deterministic HyperFrames-document.+video. Script creates stable semantic addresses; a Composition selects one+content-addressed ProgramBasis; a versioned Locator binds those addresses to that+basis; imported Track components then lower the program into one deterministic+HyperFrames document.```text.svml + .svc + .svs + .svk + lock + evidence│- ├─ Plan IR ────────> Canvas view- ├─ Located IR ─────> Timeline view- └─ HyperFrames HTML ──> rendered video+ ├─ Plan IR ───────────────> Canvas view+ ├─ TemporalBinding ───────> Timeline view+ └─ flat Track[] ──────────> HyperFrames HTML```Canvas and Timeline are views of the same source closure. They are not additional@@ -27,26 +27,25 @@ stdlib. The deterministic slice is working end to end:- closed, possibly disconnected Selection sets and zero-width Moments;- `.svc` content modules and typed `.svs` parameter classes;- stable Plan identities, separate execution digests and a verified `svml.lock`;-- isolated `.svk` projectors with scoped CSS and a JSON-only ABI;+- isolated `.svk` component lowering with scoped CSS and a JSON-only ABI;- typed `.svk` ports, parameters and recursive child-content schemas;- manifest-enforced temporal `one` / `each` / `set` consumption;-- a Speech Spine reference producer, media/presentation mapping, ranking, B-roll,+- a Speech Spine compatibility component, media/presentation mapping, ranking, B-roll,captions, text and audio;- frame-exact Located IR, Canvas/Timeline views and HyperFrames HTML.The current runtime deliberately has no live generation provider host. A-`profile="capability-v1"` Kernel remains visible in Plan, but compilation requires+`profile="capability-v1"` Component remains visible in Plan, but compilation requiresan exact, content-verified `svml.artifacts.v1` binding for its outputs and neverfalls back to a provider call. Direct existing `Image`, `Video` and `Audio` valueswork as usual. This keeps the first end-to-end reference deterministic whilepreserving the eventual capability boundary.-The executable prototype still models temporal evidence as one globally monotonic-Speech Spine alignment with shared adjacent Segment cuts. The current architecture-draft supersedes that restriction with a duck-typed `TemporalProduction` contract-and `2M + 2N` independent Segment/word endpoint identities. That migration is an-explicit implementation boundary; the repository does not claim it is already-implemented.+The executable prototype still implements frozen Script Surface v1: one globally+monotonic Speech Spine alignment with shared adjacent Segment cuts. The architecture+draft and Script Surface v2 draft specify the next boundary: a replaceable Basis+Producer, a separately versioned Locator, and `2M + 2N` independent Segment/word+endpoint identities. The repository does not claim that migration is implemented.## Commands@@ -74,20 +73,20 @@ pnpm svml compile examples/regen-ranking/regen-ranking.svml \pnpm svml render build/index.html --out build/video.mp4```-`--artifacts` is needed only when the reachable Plan contains capability Kernels.+`--artifacts` is needed only when the reachable Plan contains capability Components.`fmt` currently canonicalizes only the frozen Script body and refuses any rewritewhose reparsed semantic IR differs.`canvas` requires no timing or provider result. `estimate` deterministicallycreates provisional syllable-based alignment; `timeline` and `compile` consume-either that preview evidence or measured Speech Spine evidence. A compile is-marked reproducible only when its exact-module/Kernel/material closure matches the lock and every reachable capability-output has been content-verified.+either that preview evidence or measured Speech Spine evidence. The current+prototype verifies source closure and bound capability artifacts. It does not yet+claim complete temporal reproducibility because Evidence, Locator, ProgramBasis and+SemanticMap digests are not all part of the executable lock path.## Reference implementation[`examples/regen-ranking/regen-ranking.svml`](examples/regen-ranking/regen-ranking.svml)-reconstructs a real production video containing the Speech Spine reference producer,+reconstructs a real production video containing the Speech Spine compatibility component,a five-item ranking track, five B-roll items, captions, title, music and sound effects. The media filesare intentionally gitignored; their production provenance and frozen assertionsare recorded in@@ -98,16 +97,24 @@ The default syllable estimator predicts 36.267 seconds for the frozen 36.1-seconspeech program (0.46% duration error); it remains preview evidence, not measuredalignment.+[`examples/flat-track-launch/flat-track-launch.svml`](examples/flat-track-launch/flat-track-launch.svml)+is the readable fixture for the next architecture boundary: decoupled Basis and+Locator, Script Surface v2, continuous A-roll Present changes, B-roll transition+audio, disconnected caption selections, absolute z and Film as the sole `Track[]`+consumer. It is a draft fixture, not yet accepted by the executable v1 compiler.+## Specification- [Script Surface v1](spec/script-surface-v1.md) — the frozen, human-readablesource language for narrative, text projections, selections, and moments.+- [Script Surface v2 draft](spec/script-surface-v2-draft.md) — the syntax-preserving+ independent Segment endpoint and `2M + 2N` migration.- [Source Architecture Draft](spec/source-architecture-draft.md) — the currentdraft for `.svk` components, `.svs` parameter sheets, `.svc` content modules,- Script-dependent generation, Temporal Producers, Tracks, and a Composition root.+ Script-dependent generation, Basis Producers, flat Tracks, and a Composition root.- [Compiler prototype record](docs/compiler-prototype-v1.md) — implementedboundaries, real-video evidence and remaining work.- [External Runtime Host Architecture Draft](docs/runtime-host-architecture-draft.md)— portable runtime, Artifact, Effect/Receipt, Worker, queue, provider and- self-hosting boundaries; TemporalProduction acquisition remains+ self-hosting boundaries; TemporalBasisProduction acquisition remainsintentionally deferred.diff --git a/examples/flat-track-launch/README.md b/examples/flat-track-launch/README.mdnew file mode 100644index 00000000..0a4f5000--- /dev/null+++ b/examples/flat-track-launch/README.md@@ -0,0 +1,20 @@+# Flat Track Launch architecture fixture++This source is the readable fixture for the unimplemented Source Architecture+Draft. It deliberately exercises contracts that the current v1 compiler does not+yet accept:++- Script Surface v2 independent Segment endpoints (`2M + 2N`);+- a Basis Component that overlaps adjacent Segment media;+- a separately locked Locator that produces the ExactSemanticMap;+- one ProgramBasis and one absolute ProgramSpace;+- flat visual/audio Track contributions with no Track nesting;+- continuous A-roll source mapping across overlapping Present windows;+- B-roll transition audio inside the B-roll Track;+- disconnected caption selections, Moments and manual ProgramSpans;+- Film as the sole `Track[]` consumer.++The `@svml/std/*` imports and local assets are illustrative package paths. The+adjacent `launch.svs` makes every ProgramSpace placement explicit. Do not use this+fixture as the executable regression until the flat Track milestone lands; use+`../regen-ranking/regen-ranking.svml` for the current compiler.diff --git a/examples/flat-track-launch/flat-track-launch.svml b/examples/flat-track-launch/flat-track-launch.svmlnew file mode 100644index 00000000..5743b86f--- /dev/null+++ b/examples/flat-track-launch/flat-track-launch.svml@@ -0,0 +1,242 @@+<!--+ Architecture fixture for Source Architecture Draft.+ It is intentionally not accepted by the current Script Surface v1 prototype.+-->+<svml version="2">++ <import from="@svml/std/crossfade-speech.svk"/>+ <import from="@svml/std/media-track.svk"/>+ <import from="@svml/std/broll-track.svk"/>+ <import from="@svml/std/ranking-tier-list.svk"/>+ <import from="@svml/std/text-track.svk"/>+ <import from="@svml/std/caption-track.svk"/>+ <import from="@svml/std/audio-track.svk"/>+ <import from="@svml/std/film.svk"/>++ <import as="house" from="./launch.svs"/>++ <script version="2">+ @whole++ @prelude+ <segment id="prelude"/>+ @/prelude++ <segment id="intro">+ <HOST> @promise @emphasis Video creation is about to change @/emphasis.+ @/promise~ @ranking! @rank-svml SVML starts with meaning, not pixels.+ @/rank-svml+ </segment>++ <segment id="close">+ <HOST> @reveal One semantic source becomes @emphasis every frame @/emphasis,+ every track, and every version. @/reveal+ </segment>++ @/whole~+ </script>++ <!-- Plain values can fan out to Basis and Track Components. -->+ <video id="host-long" src="./assets/host-50s.mp4"/>+ <image id="poster" src="./assets/poster.png"/>+ <video id="product-demo" src="./assets/product-demo.mp4"/>+ <image id="rank-icon-svml" src="./assets/svml.png"/>+ <image id="rank-icon-editors" src="./assets/editors.png"/>+ <audio id="whoosh" src="./assets/whoosh.wav"/>+ <audio id="music" src="./assets/music.mp3"/>++ <!--+ Basis production assembles physical media only. It emits ProgramBasis,+ source maps, program audio and a program-bound visual facet. The locked+ Locator separately binds this Script's 2M + 2N identities to the basis.+ -->+ <crossfade-speech id="voice" overlap="500ms">+ <segment+ id="prelude"+ script={script.segment.prelude}+ source={host-long}+ sourceWindow="0s .. 10s"+ />+ <segment+ id="intro"+ script={script.segment.intro}+ source={host-long}+ sourceWindow="10s .. 30s"+ />+ <segment+ id="close"+ script={script.segment.close}+ source={host-long}+ sourceWindow="30s .. 50s"+ />+ </crossfade-speech>++ <!--+ One program-bound host facet, one continuous source mapping. Present changes+ only x/y/size/mask/z; it never restarts or seeks the video.+ -->+ <media-track id="aroll" z="200">+ <item+ id="host"+ source={voice.facets.visual}+ during={script.selection.whole}+ playback="program-map"+ class="house.host-full"+ >+ <!-- Ten seconds without words: manual projections remain legal. -->+ <present+ id="prelude-top-left"+ during={script.selection.prelude}+ window="start .. start+2.2s"+ z="210"+ class="house.face-top-left"+ exit="fade 200ms"+ />+ <present+ id="prelude-top-right"+ during={script.selection.prelude}+ window="start+2s .. start+4.2s"+ z="211"+ class="house.face-top-right"+ enter="fade 200ms"+ exit="fade 200ms"+ />+ <present+ id="prelude-bottom-right"+ during={script.selection.prelude}+ window="start+4s .. start+6.2s"+ z="212"+ class="house.face-bottom-right"+ enter="fade 200ms"+ exit="fade 200ms"+ />+ <present+ id="prelude-bottom-left"+ during={script.selection.prelude}+ window="start+6s .. start+8.2s"+ z="213"+ class="house.face-bottom-left"+ enter="fade 200ms"+ exit="fade 200ms"+ />+ <present+ id="prelude-center"+ during={script.selection.prelude}+ window="start+8s .. end"+ z="214"+ class="house.face-center"+ enter="fade 200ms"+ />++ <present+ id="promise-split"+ during={script.selection.promise}+ z="220"+ class="house.host-lower-split"+ />+ <present+ id="reveal-pip"+ during={script.selection.reveal}+ window="start-200ms .. end"+ z="221"+ class="house.host-corner-circle"+ enter="fade 200ms"+ />+ </item>+ </media-track>++ <!-- B-roll owns item composition, visual transition, source audio and SFX. -->+ <broll-track id="broll" z="100" class="house.broll-full">+ <item+ id="poster"+ source={poster}+ during={script.selection.promise}+ playback="hold"+ />+ <item+ id="product-demo"+ source={product-demo}+ during={script.selection.reveal}+ playback="stretch"+ sourceAudio="on"+ sourceGain="0.35"+ />+ <transition+ from="poster"+ to="product-demo"+ kind="crossfade"+ duration="200ms"+ sound={whoosh}+ soundGain="0.7"+ />+ </broll-track>++ <!-- Ranking is another flat Track, not a child of either media Track. -->+ <ranking-tier-list+ id="ranking"+ z="300"+ at={script.moment.ranking}+ class="house.ranking"+ >+ <item+ id="svml"+ label="SVML"+ rank="1"+ icon={rank-icon-svml}+ during={script.selection.rank-svml}+ />+ <item+ id="editors"+ label="Pixel-first editors"+ rank="2"+ icon={rank-icon-editors}+ during={script.selection.reveal}+ />+ </ranking-tier-list>++ <!-- Manual ProgramSpan is intentionally legal. -->+ <text-track id="overlays" z="400">+ <item id="opening-title" during="0s .. 3s" class="house.title">+ Semantic video starts with meaning.+ </item>+ <item id="ending-arrow" during="end-2s .. end" class="house.arrow">+ ↓+ </item>+ </text-track>++ <!-- The repeated emphasis Selection is disconnected and consumed as a set. -->+ <caption-track id="captions" z="500" script={script} class="house.captions">+ <style during={script.selection.emphasis} class="house.caption-emphasis"/>+ </caption-track>++ <audio-track id="background-music">+ <item+ id="music"+ source={music}+ during="full"+ volume="0.28"+ bus="music"+ duckUnder="voice"+ />+ </audio-track>++ <!--+ Child order is deliberately not z order. Film is the only Track[] consumer.+ Grain and master mix are Film-level because they need the complete program.+ -->+ <film+ id="launch"+ basis={voice.production}+ class="house.vertical"+ globalEffect="grain-soft"+ masterMix="house.launch-mix"+ >+ <track ref={captions.track}/>+ <track ref={background-music.track}/>+ <track ref={broll.track}/>+ <track ref={overlays.track}/>+ <track ref={aroll.track}/>+ <track ref={ranking.track}/>+ </film>++</svml>diff --git a/examples/flat-track-launch/launch.svs b/examples/flat-track-launch/launch.svsnew file mode 100644index 00000000..f90b811c--- /dev/null+++ b/examples/flat-track-launch/launch.svs@@ -0,0 +1,129 @@+<sheet version="1" id="house">++ film.vertical {+ resolution: 1080x1920;+ frame-rate: 30;+ background: #09090B;+ }++ media-track.host-full {+ x: 0;+ y: 0;+ width: 1;+ height: 1;+ fit: cover;+ }++ media-track.face-top-left {+ x: 0.05;+ y: 0.06;+ width: 0.30;+ height: 0.30;+ fit: cover;+ mask: circle;+ }++ media-track.face-top-right {+ x: 0.65;+ y: 0.06;+ width: 0.30;+ height: 0.30;+ fit: cover;+ mask: circle;+ }++ media-track.face-bottom-right {+ x: 0.65;+ y: 0.64;+ width: 0.30;+ height: 0.30;+ fit: cover;+ mask: circle;+ }++ media-track.face-bottom-left {+ x: 0.05;+ y: 0.64;+ width: 0.30;+ height: 0.30;+ fit: cover;+ mask: circle;+ }++ media-track.face-center {+ x: 0.35;+ y: 0.35;+ width: 0.30;+ height: 0.30;+ fit: cover;+ mask: circle;+ }++ media-track.host-lower-split {+ x: 0;+ y: 0.50;+ width: 1;+ height: 0.50;+ fit: cover;+ }++ media-track.host-corner-circle {+ x: 0.68;+ y: 0.70;+ width: 0.27;+ height: 0.27;+ fit: cover;+ mask: circle;+ }++ broll-track.broll-full {+ x: 0;+ y: 0;+ width: 1;+ height: 1;+ fit: cover;+ }++ ranking-tier-list.ranking {+ x: 0.06;+ y: 0.14;+ width: 0.88;+ row-gap: 0.018;+ board-color: #141418;+ accent: #8B5CF6;+ }++ text-track.title {+ x: 0.06;+ y: 0.08;+ width: 0.88;+ font-family: Inter;+ font-size: 64px;+ font-weight: 850;+ fill: #FFFFFF;+ }++ text-track.arrow {+ x: 0.42;+ y: 0.80;+ width: 0.16;+ font-size: 120px;+ fill: #A78BFA;+ }++ caption-track.captions {+ x: 0.08;+ y: 0.78;+ width: 0.84;+ font-family: Inter;+ font-size: 56px;+ font-weight: 800;+ fill: #FFFFFF;+ }++ caption-track.caption-emphasis {+ fill: #FDE047;+ scale: 1.08;+ }++</sheet>diff --git a/spec/script-surface-v1.md b/spec/script-surface-v1.mdindex e1f39898..01179c9c 100644--- a/spec/script-surface-v1.md+++ b/spec/script-surface-v1.md@@ -18,7 +18,7 @@ SVML 是语义视频源语言,不是像素渲染格式。本文只冻结它最| 构造 | 只负责 | 不负责 ||---|---|---|-| `Segment` | 有序的媒体/口播块和独立首尾锚点 | 自动 Selection、时长、静音 |+| `Segment` | 有序的媒体/口播块和结构切点 | 自动 Selection、时长、静音 || `Role Cue` | `dialogue` 导出时的“谁说了什么”文本前缀 | speaker 实体、音色、字段路由 || `Dual Text` | 显示文字与实际读音不同 | TTS 厂商参数、样式、动作 || `Selection` | 一个或多个显式闭合的语义时间区间 | 消费者声明、绝对秒数 |@@ -80,7 +80,7 @@ Meet Hypit.````silence` 只是 Selection 的名字,不会自动让音频静音。它完整选择了空-Segment 的两个独立结构锚点;真正的静音或素材行为由外部 Program 决定。+Segment 的两个结构切点;真正的静音或素材行为由外部 Program 决定。`pop` 是一个默认吸右的 Moment,解析到 `laughed` 的起音;它同样不进入任何文本投影。@@ -102,17 +102,13 @@ Segment 的两个独立结构锚点;真正的静音或素材行为由外部 Pr- Segment id 在文档内唯一,Segment 不允许嵌套。- 自然语言只能出现在 Segment 内;Segment 之间只允许空白、Selection/Moment标记和注释。-- 空 Segment 使用自闭合形式。它有独立的结构开始/结束锚点,但没有词法锚点。+- 空 Segment 使用自闭合形式。它有结构开始/结束切点,但没有词法锚点。- 空 Segment 不蕴含静音,也不蕴含任何默认时长。其物理时长来自媒体、生成结果或 `<script>` 外的 Program。- Segment 不自动声明同名 Selection。选择一个、三个或任意多个 Segment,都必须显式写闭合 Selection。-- 每个 Segment 独立拥有开始和结束锚点。因此 `m` 个 Segment 始终贡献- `2m` 个具有稳定身份的结构锚点;相邻 Segment 的两个边界即使最终落在- 同一个 ProgramPoint,也不能合并成同一身份。-- Segment 的源码顺序是叙事与作者顺序,不强迫其物理区间首尾相接。Temporal- Producer 可以把两个相邻 Segment 定位为硬切、重叠或留白,也可以让多个- 身份落到同一个点;这些关系属于定位表,不属于 Script 源语法。+- 相邻 Segment 共享一个切点。因此 `m` 个 Segment 始终贡献 `m + 1` 个+ 具有稳定身份的结构切点。例如选择三个连续 Segment:@@ -257,7 +253,7 @@ before ~@x hello @/x~ after@/pause```-`pause` 的开始吸附空 Segment 的开始锚点,结束吸附它的结束锚点。Selection+`pause` 的开始吸附空 Segment 的开始切点,结束吸附它的结束切点。Selection可完全位于 Segment 外,也可跨任意数量的 Segment。### 4.2 闭合、交叉与非连通@@ -300,53 +296,18 @@ Slot 绑定、Dual Text speech 投影和 NFC 归一化之后,v1 使用规范边界,可以在新的 timing profile / Script Surface 版本中复用相同的Selection/Moment 类型;v1 不应为尚不存在的精度引入新正文符号。-若 Script 中有 `N` 个 Segment,第 `k` 个 Segment 有 `mₖ` 个 speech token,-且总 token 数 `M = Σmₖ`,Semantic Anchor Index 恰有:+若 speech 中有 `n` 个 token、Script 中有 `m` 个 Segment,Temporal Anchor+Map 恰有:```text-Σ(2mₖ + 2) = 2M + 2N+2n + m + 1```-个稳定身份。第 `k` 个 Segment 的局部序列是:+个稳定身份:每个 token 的词首/词尾各一个,加上 `m + 1` 个唯一结构切点。+这些身份即使最终解析到同一帧也不能合并。-```text-segment[k].start-token[k,1].start-token[k,1].end-...-token[k,mₖ].start-token[k,mₖ].end-segment[k].end-```--Program 起点和终点属于 ProgramBasis,不额外进入这 `2M + 2N` 个 Script-身份。Temporal Producer 必须输出从上述每一个身份到同一 ProgramBasis 中-ProgramPoint 的完整映射;不得遗漏端点,也不得要求消费者补点。不同身份可以-落到同一点,但身份仍不能合并。--v1 默认的有序 Segment profile 只要求每个 Segment 自己内部非降序,允许任意-相邻点重合:--```text-segment.start ≤ token₁.start ≤ token₁.end ≤ ... ≤ segment.end-```--不同 Segment 之间没有全局单调、无重叠或首尾相接约束。因此:--```text-hard cut segment[a].end == segment[b].start-overlap segment[b].start < segment[a].end-gap segment[b].start > segment[a].end-```--都是合法定位结果。跨 Segment 的 Selection 仍按两个具名端点分别解析;若某个-消费者投影后得到零长或反向区间,继续遵守下文的 occurrence 丢弃规则,而不是-反向修改锚点表。--同一份 SelectionSet 可先在 estimated Complete Anchor Table 上得到预览区间,-再在 evidence/derived Complete Anchor Table 上得到最终区间。完整性与证据质量-正交:表必须始终覆盖全部身份,但每个点可以标记为 `estimated`、`derived` 或-`measured`。Script 本身不含秒数、帧号或采样点。+同一份 SelectionSet 可先在 Estimate Anchor Map 上得到预览区间,再在+Evidence Anchor Map 上得到精确区间。Script 本身不含秒数、帧号或采样点。Script Surface 不绑定帧率、采样率或渲染器。后端一旦选择物理时钟,必须只量化一次并让所有消费者复用同一整数边界;后端时钟变化不改变本语言表面。@@ -382,9 +343,8 @@ I just @pop! laughed my ass out.I laughed ~@pop! and left.```-`pop` 解析到 `laughed` 的收音。位于两个 Segment 之间时,左吸候选是前一个-Segment 的结束锚点,右吸候选是后一个 Segment 的开始锚点;两者可能落在-同一点,也可能因重叠或留白而落在不同点。+`pop` 解析到 `laughed` 的收音。位于两个 Segment 之间时,左右候选是对应+方向上的共享结构切点。规范规则:@@ -399,9 +359,9 @@ Segment 的结束锚点,右吸候选是后一个 Segment 的开始锚点;两temporal syntax,不能退化为正文。- Moment 与 Selection marker 可出现在相同位置,包括 Segment 之间和Dual Text 的 speech 侧,但均不得切入 v1 speech token。-- Moment 只选择 Complete Anchor Table 中已有的一个候选点,不新增 anchor- identity;因此 `2M + 2N` 的精确计数不因 Moment 数量改变。-- MomentSet 与 SelectionSet 复用同一份 Complete Anchor Table 和+- Moment 只选择 Temporal Anchor Map 中已有的一个候选点,不新增 anchor+ identity;因此 `2n + m + 1` 的精确计数不因 Moment 数量改变。+- MomentSet 与 SelectionSet 复用同一份 Estimate/Evidence Anchor Map 和一次性帧量化;区别只在最终载体是 `Point[]` 而不是 `Range[]`。消费端保持两个互不相混的类型:@@ -503,7 +463,7 @@ temporal marker、Role Cue 或结构标签内部。注释可出现在 atom 之解析结果至少保留:-- Segment 的顺序、id、结构边界和各自独立的 start/end anchor identity;+- Segment 的顺序、id 和结构边界;- 每个 Segment 的有序 spoken turn,以及各 turn 的可选 Role Cue;- plain / Dual Text / Slot atom 及其源码范围;- caption atom 到一个或多个 speech token 的显式映射;@@ -543,7 +503,7 @@ Caption 的多样式、region、cue segmentation、annotation、mute 和 layout| 绝对秒数、帧号、采样点、任意 sub-frame 位置 | Script 外的运行时/Program || token 内音素或字素级切点、标点专属时间 | 后续 timing profile 或新版本 || 空 Segment 的预览时长 | 媒体、生成器或显式 Program |-| 同时重叠说话 | 不新增正文语法;Temporal Producer 可把不同 Segment 定位为重叠,Caption/Track 分别声明自己的 overlap policy |+| 同时重叠说话 | 多轨媒体/生成 Program;Base Script 仍是顺序语义轴 || 笑声、咳嗽、环境声、SFX、beep | 音频/生成/SFX Program || 无 speech 时间依据的纯显示文字 | Text、Deck 或 Caption Program || 字幕样式、手工断行、动效、画面动作 | 相应 Program 引用 Selection |@@ -577,10 +537,7 @@ Caption 的多样式、region、cue segmentation、annotation、mute 和 layout2. parser、formatter、Narrative IR 与三投影不依赖 provider 或渲染器;3. SelectionSet 支持闭合、交叉、非连通 occurrence 与左右 affinity;4. MomentSet 支持 `@id!` / `~@id!`、单/多 occurrence 与 `at` 类型闸;-5. Semantic Anchor Index 恰有 `2M + 2N` 个独立身份;Complete Anchor Table- 覆盖硬切、重叠、留白、空 Segment、端点重合和跨 Segment Selection,并且- 只逐 Segment 校验单调性;-6. Dual Text 在 speech、caption 与时间映射上保持一个可审计 source map;-7. Slot 采用 parse-first、literal-only binding,不能注入语法;-8. 未知版本、未知保留语法、混合 temporal type 与 partial dual atom 全部+5. Dual Text 在 speech、caption 与时间映射上保持一个可审计 source map;+6. Slot 采用 parse-first、literal-only binding,不能注入语法;+7. 未知版本、未知保留语法、混合 temporal type 与 partial dual atom 全部fail closed。