@hypit/hypit 0.1.14 → 0.2.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (82) hide show
  1. package/README.md +2 -2
  2. package/dist/public/browser-capture.d.ts +9 -3
  3. package/dist/public/caption.d.ts +2 -0
  4. package/dist/public/narrative.d.ts +2 -0
  5. package/dist/public/runtime-kit.d.ts +2 -2
  6. package/dist/public/speech.d.ts +2 -0
  7. package/examples/semantic-composition/packages/responsive-explainer/README.md +1 -1
  8. package/package.json +1 -1
  9. package/packages/browser-capture/README.md +9 -2
  10. package/packages/browser-capture/package.json +5 -0
  11. package/packages/browser-capture/src/browser.ts +22 -8
  12. package/packages/browser-capture/src/index.ts +8 -2
  13. package/packages/caption/README.md +1 -1
  14. package/packages/caption/src/display.ts +1 -0
  15. package/packages/caption-fine/README.md +19 -5
  16. package/packages/caption-fine/package.json +1 -0
  17. package/packages/caption-fine/src/joined-box.ts +114 -0
  18. package/packages/caption-fine/src/manifest.ts +1 -1
  19. package/packages/caption-fine/src/render.ts +99 -92
  20. package/packages/caption-fine/src/spacing.ts +8 -61
  21. package/packages/caption-fine-studio/src/index.ts +2 -4
  22. package/packages/cli/README.md +3 -2
  23. package/packages/cli/src/arguments.ts +2 -2
  24. package/packages/cli/src/command.ts +1 -1
  25. package/packages/cli/src/commands/environment.ts +9 -2
  26. package/packages/cli/src/machine-view.ts +1 -1
  27. package/packages/cli/src/output.ts +1 -0
  28. package/packages/credential-store-os/src/store.ts +5 -1
  29. package/packages/narrative/README.md +3 -1
  30. package/packages/narrative/src/schema.ts +1 -0
  31. package/packages/narrative/src/types.ts +2 -0
  32. package/packages/provider-whisperx-local/README.md +26 -10
  33. package/packages/provider-whisperx-local/src/activation.ts +10 -1
  34. package/packages/provider-whisperx-local/src/program.ts +8 -4
  35. package/packages/runtime-host-node/src/index.ts +1 -0
  36. package/packages/runtime-kit/README.md +6 -2
  37. package/packages/runtime-kit/src/index.ts +2 -2
  38. package/packages/runtime-local/README.md +6 -1
  39. package/packages/runtime-local/src/host.ts +2 -0
  40. package/packages/runtime-local/src/index.ts +1 -0
  41. package/packages/runtime-local/src/programs.ts +53 -17
  42. package/packages/script/README.md +81 -34
  43. package/packages/script/bin/migrate-0.2.mjs +90 -0
  44. package/packages/script/src/edit.ts +28 -60
  45. package/packages/script/src/lexical.ts +59 -130
  46. package/packages/script/src/manifest.ts +11 -11
  47. package/packages/script/src/narrative.ts +4 -2
  48. package/packages/script/src/parser.ts +140 -193
  49. package/packages/script/src/types.ts +2 -3
  50. package/packages/studio/src/ui/code.ts +1 -1
  51. package/packages/studio/src/ui/main.ts +2 -2
  52. package/packages/studio/src/ui/markers.ts +3 -3
  53. package/packages/studio/src/ui/selection.ts +1 -1
  54. package/packages/studio/src/ui/syntax.ts +42 -21
  55. package/packages/temporal-markup/README.md +1 -1
  56. package/packages/video-cli/README.md +9 -5
  57. package/packages/video-cli/src/capture.ts +17 -2
  58. package/packages/video-cli/src/creation.ts +4 -6
  59. package/packages/video-cli/src/media.ts +10 -3
  60. package/packages/whisperx/README.md +56 -3
  61. package/packages/whisperx/src/component.ts +2 -5
  62. package/packages/whisperx/src/evidence.ts +4 -3
  63. package/packages/whisperx/src/index.ts +1 -0
  64. package/packages/whisperx/src/manifest.ts +2 -3
  65. package/packages/whisperx/src/surface.ts +3 -6
  66. package/packages/whisperx/src/types.ts +9 -1
  67. package/packages/yt-dlp/README.md +9 -3
  68. package/packages/yt-dlp/package.json +2 -1
  69. package/packages/yt-dlp/src/download.ts +10 -24
  70. package/packages/yt-dlp/src/environment.ts +36 -0
  71. package/packages/yt-dlp/src/index.ts +1 -0
  72. package/services/whisperx/README.md +45 -12
  73. package/services/whisperx/src/hypit_whisperx_service/application.py +3 -0
  74. package/services/whisperx/src/hypit_whisperx_service/check.py +7 -0
  75. package/services/whisperx/src/hypit_whisperx_service/config.py +7 -0
  76. package/services/whisperx/src/hypit_whisperx_service/engine.py +21 -5
  77. package/services/whisperx/src/hypit_whisperx_service/models.py +131 -0
  78. package/services/whisperx/src/hypit_whisperx_service/prepare.py +21 -3
  79. package/services/whisperx/src/hypit_whisperx_service/resources.py +19 -3
  80. package/services/yt-dlp/README.md +14 -15
  81. package/services/yt-dlp/pyproject.toml +1 -1
  82. package/services/yt-dlp/uv.lock +513 -2
@@ -1,3 +1,4 @@
1
+ import { resolve } from "node:path";
1
2
  import {
2
3
  createRuntimeEndpointAdapterFacet,
3
4
  runtimeConfigExact,
@@ -18,7 +19,7 @@ const localWhisperXRuntimeAdapter = createRuntimeEndpointAdapterFacet({
18
19
  "baseUrl", "expectedModel", "expectedDevice", "expectedCompute", "expectedBatchSize",
19
20
  "expectedServiceVersion", "expectedWhisperXVersion",
20
21
  "defaultConcurrency", "requestTimeoutMs", "maxResponseBytes",
21
- "serviceCommand",
22
+ "serviceCommand", "alignmentLanguages", "modelCacheDirectory",
22
23
  ], "local WhisperX");
23
24
  const baseUrl = runtimeConfigString(config.baseUrl, "WhisperX baseUrl");
24
25
  if (baseUrl !== undefined) {
@@ -37,6 +38,12 @@ const localWhisperXRuntimeAdapter = createRuntimeEndpointAdapterFacet({
37
38
  const defaultConcurrency = runtimeConfigPositiveInteger(config.defaultConcurrency, "WhisperX defaultConcurrency");
38
39
  const requestTimeoutMs = runtimeConfigPositiveInteger(config.requestTimeoutMs, "WhisperX requestTimeoutMs");
39
40
  const maxResponseBytes = runtimeConfigPositiveInteger(config.maxResponseBytes, "WhisperX maxResponseBytes");
41
+ const modelCacheDirectory = runtimeConfigString(config.modelCacheDirectory, "WhisperX modelCacheDirectory");
42
+ const alignmentLanguages = config.alignmentLanguages;
43
+ if (alignmentLanguages !== undefined && (!Array.isArray(alignmentLanguages)
44
+ || alignmentLanguages.some((item) => typeof item !== "string" || !/^[a-z]+$/u.test(item)))) {
45
+ throw new Error("WhisperX alignmentLanguages must be an array of lowercase language codes");
46
+ }
40
47
  const serviceCommandValue = config.serviceCommand;
41
48
  if (serviceCommandValue !== undefined
42
49
  && (!Array.isArray(serviceCommandValue) || serviceCommandValue.length === 0
@@ -80,6 +87,8 @@ const localWhisperXRuntimeAdapter = createRuntimeEndpointAdapterFacet({
80
87
  expectedServiceVersion: selectedServiceVersion,
81
88
  expectedWhisperXVersion: selectedWhisperXVersion,
82
89
  ...(serviceCommand === undefined ? {} : { serviceCommand }),
90
+ ...(alignmentLanguages === undefined ? {} : { alignmentLanguages: alignmentLanguages as string[] }),
91
+ ...(modelCacheDirectory === undefined ? {} : { modelCacheDirectory: resolve(context.dataRoot, modelCacheDirectory) }),
83
92
  }),
84
93
  };
85
94
  },
@@ -52,6 +52,8 @@ export type LocalWhisperXProgramOptions = {
52
52
  readonly expectedServiceVersion: string;
53
53
  readonly expectedWhisperXVersion: string;
54
54
  readonly serviceCommand?: ManagedProgramCommand;
55
+ readonly alignmentLanguages?: readonly string[];
56
+ readonly modelCacheDirectory?: string;
55
57
  };
56
58
 
57
59
  export function localWhisperXProgram(options: LocalWhisperXProgramOptions): ManagedProgram {
@@ -85,11 +87,13 @@ export function localWhisperXProgram(options: LocalWhisperXProgramOptions): Mana
85
87
  HYPIT_WHISPERX_COMPUTE: expected.compute,
86
88
  HYPIT_WHISPERX_BATCH_SIZE: String(expected.batchSize),
87
89
  HYPIT_WHISPERX_NLTK_DATA: nltkData,
90
+ ...(options.modelCacheDirectory === undefined ? {} : { HYPIT_WHISPERX_MODEL_CACHE: options.modelCacheDirectory }),
91
+ HYPIT_WHISPERX_ALIGNMENT_LANGUAGES: (options.alignmentLanguages ?? []).join(" "),
88
92
  };
89
93
  const check: ManagedProgramCommand = {
90
94
  command: pythonEnvironmentCommand(environment, "hypit-whisperx-check"),
91
- args: [],
92
- env: { HYPIT_WHISPERX_NLTK_DATA: nltkData },
95
+ args: ["--models"],
96
+ env: serviceEnvironment,
93
97
  };
94
98
  const managedStart: ManagedProgramCommand = {
95
99
  command: pythonEnvironmentCommand(environment, "hypit-whisperx-service"),
@@ -131,10 +135,10 @@ export function localWhisperXProgram(options: LocalWhisperXProgramOptions): Mana
131
135
  args: ["sync", "--project", localWhisperXManagedProject, "--frozen", "--no-editable"],
132
136
  env: { UV_PROJECT_ENVIRONMENT: environment },
133
137
  }, {
134
- label: "Prepare NLTK sentence data",
138
+ label: `Prepare WhisperX ${expected.model} and alignment resources for ${(options.alignmentLanguages ?? []).join(", ") || "no selected languages"}`,
135
139
  command: pythonEnvironmentCommand(environment, "hypit-whisperx-prepare"),
136
140
  args: ["--nltk-data", nltkData],
137
- env: { HYPIT_WHISPERX_NLTK_DATA: nltkData },
141
+ env: serviceEnvironment,
138
142
  }],
139
143
  },
140
144
  } : {}),
@@ -245,6 +245,7 @@ export type RuntimeController = {
245
245
  down(options?: { readonly maxWaitMs?: number }): Promise<RuntimeWorkerState>;
246
246
  };
247
247
  readonly programs: {
248
+ prepare(options?: { readonly endpoints?: readonly string[]; readonly onProgress?: (event: ManagedProgramProgress) => void }): Promise<{ readonly dataRoot: string; readonly programs: readonly ManagedProgramReport[] }>;
248
249
  up(options: {
249
250
  readonly maxWaitMs?: number;
250
251
  readonly onProgress?: (event: ManagedProgramProgress) => void;
@@ -43,8 +43,12 @@ video package by name.
43
43
 
44
44
  `ManagedProgram.installation.prepareBeforeStart` lets a Provider reconcile its installed environment
45
45
  before a cold start, even when the installation probe already passes. The declared commands use the
46
- Provider's ordinary package manager. A healthy running Program is reused before this preparation is
47
- considered; Runtime does not inspect source files or infer implementation versions.
46
+ Provider's ordinary package manager. `programs prepare` runs the declared installation independently
47
+ of process startup; `programs up` also checks its resources when the process is already healthy.
48
+ The installation probe covers the owner's selected execution resources, while the process probe
49
+ reports service readiness. Cold reconciliation still applies only before starting a new process.
50
+ Runtime does not name models, inspect source files or infer implementation versions. Preparation
51
+ may acquire resources; execution must consume already prepared resources and report what is missing.
48
52
  `ManagedProgramCommand.label` optionally names a command's purpose for progress and log headings.
49
53
  It is display text supplied by the Program owner, not a phase to persist or interpret. Runtime
50
54
  reports generic process/probe facts; the service itself owns domain-specific progress in its logs.
@@ -42,11 +42,11 @@ export type ManagedProgramCommand = {
42
42
  };
43
43
 
44
44
  export type ManagedProgramInstallation = {
45
- /** Truthful inspection of the installed program; no installer-owned receipt. */
45
+ /** Read-only inspection of the selected environment and resources; no downloads or receipts. */
46
46
  probe(): Promise<ManagedProgramState>;
47
47
  /** Reconcile an installed environment before a cold start, using its package manager's cache. */
48
48
  readonly prepareBeforeStart?: boolean;
49
- /** Commands run when installation is missing, or before a cold start when requested above. */
49
+ /** Explicit preparation commands, run for missing resources or cold reconciliation above. */
50
50
  readonly commands: readonly ManagedProgramCommand[];
51
51
  };
52
52
 
@@ -242,7 +242,12 @@ controller returned by `await host.controller()`, and calls `build(...)` on the
242
242
  `await host.createRuntime()`. There is one production process-lifecycle path.
243
243
 
244
244
  Managed Programs have independent lifetimes. A healthy WhisperX service keeps its model loaded;
245
- `prepareBeforeStart` reconciles a cold installation through uv's source-aware synchronization.
245
+ `programs prepare --endpoint <instance>` prepares selected resources without starting or stopping
246
+ that service. Its installation probe is independent of process health, so an online service does not
247
+ hide a newly requested language model. `programs up` prepares missing resources and starts the service
248
+ if needed. `prepareBeforeStart` additionally reconciles a cold installation through the package
249
+ manager's source-aware synchronization. The Provider owns the resource choices and commands;
250
+ Runtime only invokes them under the existing lifecycle lock.
246
251
  `runtime down` stops the coordinator and asks its executors to stop; Programs are stopped separately.
247
252
  Distribution changes and shell environment changes still concern the coordinator's bootstrap process.
248
253
  Inspect active work before restarting it. Environment-backed credentials use that inherited process
@@ -23,6 +23,7 @@ import {
23
23
  } from "./config.js";
24
24
  import {
25
25
  bringManagedProgramsUp,
26
+ prepareManagedPrograms,
26
27
  reportManagedPrograms,
27
28
  takeManagedProgramsDown,
28
29
  } from "./programs.js";
@@ -82,6 +83,7 @@ export async function openLocalRuntimeHost(
82
83
  ),
83
84
  },
84
85
  programs: {
86
+ prepare: async (programOptions) => await prepareManagedPrograms(profile, { ...programOptions, packageRoot, ...distribution }),
85
87
  up: async (programOptions) => await bringManagedProgramsUp(profile, { ...programOptions, packageRoot, ...distribution }),
86
88
  down: async (scope) => await takeManagedProgramsDown(profile, { ...scope, packageRoot, ...distribution }),
87
89
  report: async (scope) => await reportManagedPrograms(profile, { ...scope, packageRoot, ...distribution }),
@@ -28,6 +28,7 @@ export {
28
28
  } from "./config.js";
29
29
  export {
30
30
  bringManagedProgramsUp,
31
+ prepareManagedPrograms,
31
32
  reportManagedPrograms,
32
33
  takeManagedProgramsDown,
33
34
  } from "./programs.js";
@@ -293,6 +293,8 @@ async function bringUpOwned(
293
293
  onProgress?.({ id: program.id, phase: "checking" });
294
294
  const initial = await program.probe();
295
295
  if (initial.state === "ready") {
296
+ const prepared = await prepareOwned(root, program, endpoint, onProgress);
297
+ if (prepared.state.state !== "ready") return prepared;
296
298
  onProgress?.({ id: program.id, phase: "ready" });
297
299
  return { ...base, action: "already-running", state: initial };
298
300
  }
@@ -306,6 +308,10 @@ async function bringUpOwned(
306
308
  const installLogPath = join(directory(root, program), "install.log");
307
309
  const existingPid = await readPid(root, program);
308
310
  if (existingPid !== undefined && processAlive(existingPid)) {
311
+ if (program.installation !== undefined) {
312
+ const prepared = await prepareOwned(root, program, endpoint, onProgress);
313
+ if (prepared.state.state !== "ready") return prepared;
314
+ }
309
315
  release();
310
316
  onProgress?.({ id: program.id, phase: "waiting", logPath,
311
317
  detail: `observing existing process ${existingPid}` });
@@ -320,23 +326,9 @@ async function bringUpOwned(
320
326
  }
321
327
  if (existingPid !== undefined) await rm(join(directory(root, program), "process.pid"), { force: true });
322
328
  if (program.installation !== undefined) {
323
- const installation = await program.installation.probe();
324
- if (installation.state !== "ready" || program.installation.prepareBeforeStart === true) {
325
- await mkdir(directory(root, program), { recursive: true });
326
- await rotateLog(installLogPath);
327
- for (const command of program.installation.commands) {
328
- onProgress?.({ id: program.id, phase: "installing", logPath: installLogPath, detail: commandLabel(command) });
329
- const result = await run(root, command, installLogPath);
330
- if (!result.ok) {
331
- return { ...base, action: "unchanged", state: initial, logPath: installLogPath, detail: `${command.command} failed: ${result.detail}` };
332
- }
333
- }
334
- const after = await program.installation.probe();
335
- if (after.state !== "ready") {
336
- return { ...base, action: "unchanged", state: initial, logPath: installLogPath, detail: `installation is ${after.state}: ${after.detail}` };
337
- }
338
- installed = true;
339
- }
329
+ const prepared = await prepareOwned(root, program, endpoint, onProgress, program.installation.prepareBeforeStart === true);
330
+ if (prepared.state.state !== "ready") return prepared;
331
+ installed = prepared.action === "installed";
340
332
  }
341
333
  if (program.start === undefined) {
342
334
  // Nothing to keep running: installation was the whole job.
@@ -435,6 +427,50 @@ async function bringUp(
435
427
  } finally { release(); }
436
428
  }
437
429
 
430
+ /** Resource preparation is independent of process readiness. Caller owns the lifecycle lock. */
431
+ async function prepareOwned(
432
+ root: string, program: ManagedProgram, endpoint: string,
433
+ onProgress?: (event: ManagedProgramProgress) => void, reconcile = false,
434
+ ): Promise<ManagedProgramReport> {
435
+ const base = { id: program.id, endpoint };
436
+ const installation = program.installation;
437
+ if (installation === undefined) return { ...base, action: "unchanged", state: await program.probe() };
438
+ const logPath = join(directory(root, program), "install.log");
439
+ const before = await installation.probe();
440
+ if (before.state === "ready" && !reconcile) return { ...base, action: "unchanged", state: before };
441
+ await mkdir(root, { recursive: true });
442
+ await rotateLog(logPath);
443
+ for (const command of installation.commands) {
444
+ onProgress?.({ id: program.id, phase: "installing", logPath, detail: commandLabel(command) });
445
+ const result = await run(root, command, logPath);
446
+ if (!result.ok) return { ...base, action: "unchanged", state: { state: "down", detail: `${command.command} failed: ${result.detail}` }, installationLogPath: logPath };
447
+ }
448
+ const state = await installation.probe();
449
+ return { ...base, action: state.state === "ready" ? "installed" : "unchanged", state, installationLogPath: logPath };
450
+ }
451
+
452
+ /** Prepare selected resources without starting, stopping or trusting a running process. */
453
+ export async function prepareManagedPrograms(
454
+ path: string, options: ManagedProgramOptions = {},
455
+ ): Promise<{ readonly dataRoot: string; readonly programs: readonly ManagedProgramReport[] }> {
456
+ const { dataRoot, programs } = await declaredManagedPrograms(path, options);
457
+ const reports = await Promise.all(independentPrograms(programs).map(async ({ instance, program }): Promise<ManagedProgramReport> => {
458
+ const home = directory(dataRoot, program);
459
+ await mkdir(home, { recursive: true });
460
+ const { tryProgramLock } = await import("./program-lock.js");
461
+ const release = tryProgramLock(home);
462
+ if (release === undefined) return { id: program.id, endpoint: instance, action: "unchanged", state: { state: "down", detail: "another command owns this Program's preparation; inspect its logs" } };
463
+ try {
464
+ options.onProgress?.({ id: program.id, phase: "checking" });
465
+ const report = await prepareOwned(dataRoot, program, instance, options.onProgress);
466
+ if (report.state.state === "ready") options.onProgress?.({ id: program.id, phase: "ready" });
467
+ return report;
468
+ }
469
+ finally { release(); }
470
+ }));
471
+ return { dataRoot, programs: reports };
472
+ }
473
+
438
474
  /** Install when needed and start every external program the Runtime Profile implies. */
439
475
  export async function bringManagedProgramsUp(
440
476
  path: string,
@@ -10,7 +10,7 @@ know that Script, Segment or Narrative exist.
10
10
 
11
11
  ```svml
12
12
  <script id="story">
13
- @answer
13
+ @{answer}
14
14
 
15
15
  <opening>
16
16
  <ALICE> I will make the first point.
@@ -18,7 +18,7 @@ know that Script, Segment or Narrative exist.
18
18
  </opening>
19
19
 
20
20
  <pause/>
21
- @/answer
21
+ @{/answer}
22
22
  </script>
23
23
  ```
24
24
 
@@ -65,8 +65,8 @@ CaptionDocument contracts.
65
65
  Role Cue or the Segment close ends that turn.
66
66
  - **Dual Text**: one authored speech span with separate display and spoken projections, written
67
67
  `<display text | spoken text>`. In `<display text|>`, omitted speech inherits the displayed prose.
68
- - **Selection marker**: a named semantic range, written `@name ... @/name`.
69
- - **Moment marker**: a named semantic point, written `@name!`.
68
+ - **Selection marker**: a named semantic range, written `@{name} ... @{/name}`.
69
+ - **Moment marker**: a named semantic point, written `@{name!}`.
70
70
  - **CaptionDocument**: the Script-owned caption truth; it contains **Display Words**,
71
71
  **Alignment Units** and **Cue Breaks**. It contains no seconds or frames.
72
72
  - **Token attribute**: a flat postfix display-word annotation such as `really{emphasis}` or
@@ -86,7 +86,7 @@ Caption Styles choose whether to highlight, reveal or keep the text steady.
86
86
 
87
87
  In a shared side, markers and display attributes do not enter either text projection. Its speech
88
88
  Tokens retain offsets into the actually written left-hand text, so Studio can move an anchor inside
89
- `<组@beat!件化|>` without expanding the shorthand. Attributes still annotate the preceding display
89
+ `<组@{beat!}件化|>` without expanding the shorthand. Attributes still annotate the preceding display
90
90
  word, not the whole group. Whitespace-only speech is omitted; a group with no spoken text on either
91
91
  side is invalid. This adds no new public value type or protocol version.
92
92
 
@@ -101,8 +101,8 @@ Units; it does not split the Segment, cut the picture or end a Selection.
101
101
  <script id="story">
102
102
  <exchange>
103
103
  <HOST> I use it || every day, || since <2012 | twenty twelve>.
104
- <GUEST> Even @proof on holiday @/proof?
105
- <HOST> @answer! Especially then.
104
+ <GUEST> Even @{proof} on holiday @{/proof}?
105
+ <HOST> @{answer!} Especially then.
106
106
  </exchange>
107
107
  </script>
108
108
  ```
@@ -117,19 +117,53 @@ following characters. Display punctuation attaches to neighboring words without
117
117
  These units support precise timing and highlighting. A Caption Cue can hold a whole phrase of them;
118
118
  `||` chooses its handoff independently of character counts or visual line wrapping.
119
119
 
120
+ Annotations do not create speech boundaries. Comments are transparent (`hel<!--note-->lo`
121
+ remains `hello`); a postfix attribute or `||` inside a word is invalid. Script analyzes a complete
122
+ prose run before binding these constructs. An explicit Dual correspondence and a speaker/Segment
123
+ boundary remain authored structure. Shared Dual groups still expose their internal speech anchors.
124
+
125
+ A Dual display side is literal authored text, including symbols and emoji: `<😀|smile>` and
126
+ `<.|dot>` have explicit speech correspondence and require no invented speech token for the symbol.
127
+ A literal-only display is one display surface within that correspondence.
128
+
129
+ ## Display spelling and separators
130
+
131
+ Script preserves the normalized display spelling independently of speech tokenization. Ordinary
132
+ whitespace runs become one space; leading/trailing whitespace in a Turn and padding at the edges
133
+ of a Dual Text side are omitted. No language-specific rule removes a Chinese space or inserts a
134
+ space between numeric and Korean/Latin tokens. `是的 就是这样`, `3개월`, `3 개월`, `3D` and `3 D`
135
+ therefore remain distinct as authored. Source newlines are prose formatting, not Caption Cue breaks.
136
+ Use `||` for Cues and a family's layout controls for visual rows.
137
+
138
+ Each `CaptionDisplayWord.separatorBefore` is `""` or `" "`, relative to the preceding displayed word
139
+ in its Turn. Together with `text`, it carries the display spelling to consumers; it is not a speech
140
+ Token and has no timing. A renderer suppresses the leading separator at a displayed line/Cue start.
141
+ Shared groups retain internal separators and individual speech anchors: `<New York|>` is one
142
+ Alignment Unit with a space inside, while `<3D|>` has none. Grouping is a creative choice, not a
143
+ workaround for preserving spelling.
144
+
120
145
  ## Selection and Moment affinity
121
146
 
147
+ Every marker starts with `@{` and ends with `}`. The complete marker is zero-width in speech and
148
+ display; surrounding prose whitespace remains prose. All control sigils belong inside the braces:
149
+ `@{beat!}` is a Moment, whereas `@{part}!` opens a Selection followed by a literal exclamation mark.
150
+ Names match `[a-z][a-z0-9_-]{0,63}`; whitespace and nesting inside a marker are invalid.
151
+ A marker cannot split a speech Token or separate it from attached punctuation: place
152
+ `@{beat!}“测试”`, not `“@{beat!}测试”`. `hello{emphasis}` is a postfix display attribute; `@{part}`
153
+ is consumed as one marker and cannot be mistaken for that attribute. Write a literal `@{part}` as
154
+ `\@\{part\}`.
155
+
122
156
  A marker selects the adjacent semantic anchor; it does not write a timecode. Inside a spoken
123
157
  passage, its affinity normally chooses a neighboring word boundary:
124
158
 
125
159
  | Marker | Boundary |
126
160
  | --- | --- |
127
- | `@name` | Selection opens at the next word's start: right affinity. |
128
- | `~@name` | Selection opens at the previous word's end: left affinity. |
129
- | `@/name` | Selection closes at the previous word's end: left affinity. |
130
- | `@/name~` | Selection closes at the next word's start: right affinity. |
131
- | `@name!` | Moment at the next word's start: right affinity. |
132
- | `~@name!` | Moment at the previous word's end: left affinity. |
161
+ | `@{name}` | Selection opens at the next word's start: right affinity. |
162
+ | `@{~name}` | Selection opens at the previous word's end: left affinity. |
163
+ | `@{/name}` | Selection closes at the previous word's end: left affinity. |
164
+ | `@{/name~}` | Selection closes at the next word's start: right affinity. |
165
+ | `@{name!}` | Moment at the next word's start: right affinity. |
166
+ | `@{~name!}` | Moment at the previous word's end: left affinity. |
133
167
 
134
168
  At structural edges, the parser retains the corresponding Segment or program boundary rather than
135
169
  inventing a neighboring word. Selection and Moment ids share one namespace. Selections can overlap,
@@ -138,37 +172,50 @@ cross and span Segments; unlike tags, they do not need to nest.
138
172
  To join two visual Selections without exposing their inter-word pause, match affinity on both sides:
139
173
 
140
174
  ```text
141
- @coffee my coffee @/coffee ~@smoothie my smoothie @/smoothie
142
- @coffee my coffee @/coffee~ @smoothie my smoothie @/smoothie
175
+ @{coffee} my coffee @{/coffee} @{~smoothie} my smoothie @{/smoothie}
176
+ @{coffee} my coffee @{/coffee~} @{smoothie} my smoothie @{/smoothie}
143
177
  ```
144
178
 
145
179
  These are alternative spellings, not two occurrences to put in the same Script. In the first,
146
180
  “coffee” ends both the first Window and the gap's left boundary, so the smoothie Selection owns the
147
181
  pause. In the second, the next “my” starts both touching boundaries, so the coffee Selection owns it.
148
- Plain `@/coffee @smoothie` leaves the gap between previous word end and next word start outside both.
182
+ Plain `@{/coffee} @{smoothie}` leaves the gap between previous word end and next word start outside both.
149
183
  The media consumer still decides playback and visual coverage inside those projected Windows.
150
184
 
151
185
  ## Marker writeback
152
186
 
153
187
  `adjustScriptSelection` and `adjustScriptMoment` accept explicit anchor identities. A Selection's
154
- two endpoints are written together. Script owns their legal source sites: token punctuation and
155
- postfix attributes stay attached, Segment boundaries stay structural, and coincident markers are
156
- ordered together. Writeback reads the current source, removes the markers, normalizes ordinary
157
- same-line prose, and places the markers at their requested anchors. Repeated spaces and tabs in
158
- prose become necessary separators; punctuation uses the same attachment rules as Script's text
159
- projection. Newlines, blank lines and each line's leading spaces/tabs remain intact. A marker at
160
- the start of a line is placed after its indentation; whitespace after that marker is prose, not
161
- additional indentation. Marker-only lines remain blank when the marker leaves. Comments, tags,
162
- Dual Text and display attributes retain their own syntax and are not passed through prose cleanup.
163
-
164
- Normalization keeps lexical units unchanged and does not restore earlier whitespace spellings.
165
- It has no persistent formatting state: the same text structure and anchor relationships produce
166
- the same spelling on repeated edits. Empty self-closing Segments expand when needed to make their
167
- distinct boundaries writable.
168
-
169
- Writeback reparses the result to retain the intended identities and unchanged narrative/caption
170
- content. These checks concern authored order, not frame order. A consumer's temporal projection
171
- determines whether the resulting Instant or Window is usable.
188
+ two endpoints are moved together; unrelated markers and all source prose remain untouched. Names
189
+ terminate at `}`, so writeback never inserts a separating space or normalizes surrounding prose.
190
+ A no-op adjustment returns the original source. Script owns legal insertion sites: Tokens and
191
+ postfix attributes stay together, Dual Text markers follow the actual speech side, and coincident
192
+ markers written together have deterministic order. Empty self-closing Segments expand only when a
193
+ requested boundary needs an interior insertion site.
194
+
195
+ Writeback reparses the result to retain the requested bindings and unchanged speech, display and
196
+ Narrative/Caption content. Source offsets remain parser-private; no formatting history is stored.
197
+
198
+ ## Explicit migration from 0.1
199
+
200
+ The 0.2 parser rejects bare `@name` markers. Preview migration from the repository or installed
201
+ Distribution root, then explicitly write the reviewed result:
202
+
203
+ ```sh
204
+ node packages/script/bin/migrate-0.2.mjs /path/to/film.svml
205
+ node packages/script/bin/migrate-0.2.mjs /path/to/film.svml --write
206
+ ```
207
+
208
+ Use `--body` for a file containing a raw Script body rather than outer SVML. The tool converts
209
+ markers only inside Script bodies, leaves comments and escapes intact, and does not touch provider
210
+ prompt references such as `@image1`. It neither installs anything nor runs during a build.
211
+
212
+ The tool changes marker spelling, not marker placement. Move a marker that separates a word from
213
+ its attached quote or punctuation to the complete word boundary before using that source.
214
+
215
+ Review authored whitespace after migration: spaces previously discarded by Chinese/punctuation
216
+ normalization now appear. The tool preserves source spaces rather than guessing the author's intent.
217
+ Regenerate affected Narrative, caption and Build results with the new reader/writer together;
218
+ protocol identities remain `@1`. Existing rendered media is not modified by source migration.
172
219
 
173
220
  ## Complete authored content and narrow exports
174
221
 
@@ -0,0 +1,90 @@
1
+ #!/usr/bin/env node
2
+ /** Explicit, one-time marker migration. Never called by parsing, building, or Studio. */
3
+ import { readFile, writeFile } from "node:fs/promises";
4
+ import { resolve } from "node:path";
5
+ import { pathToFileURL } from "node:url";
6
+
7
+ export function migrateScriptBody(source) {
8
+ let result = "";
9
+ for (let cursor = 0; cursor < source.length;) {
10
+ if (source[cursor] === "\\") {
11
+ result += source.slice(cursor, cursor + 2);
12
+ cursor += 2;
13
+ continue;
14
+ }
15
+ if (source.startsWith("<!--", cursor)) {
16
+ const close = source.indexOf("-->", cursor + 4);
17
+ if (close < 0) throw new Error("Unclosed Script comment");
18
+ result += source.slice(cursor, close + 3);
19
+ cursor = close + 3;
20
+ continue;
21
+ }
22
+ if (source.startsWith("@{", cursor)) {
23
+ const close = source.indexOf("}", cursor + 2);
24
+ if (close < 0) throw new Error("Unclosed delimited Script marker");
25
+ result += source.slice(cursor, close + 1);
26
+ cursor = close + 1;
27
+ continue;
28
+ }
29
+ const rest = source.slice(cursor);
30
+ const match = /^@\/([a-z][a-z0-9_-]{0,63})(~)/u.exec(rest)
31
+ ?? /^@\/([a-z][a-z0-9_-]{0,63})(?![A-Za-z0-9_-])/u.exec(rest);
32
+ const open = match ? undefined : /^(~)?@([a-z][a-z0-9_-]{0,63})(!)/u.exec(rest)
33
+ ?? /^(~)?@([a-z][a-z0-9_-]{0,63})(?![A-Za-z0-9_-])/u.exec(rest);
34
+ if (match) {
35
+ result += `@{/${match[1]}${match[2] ?? ""}}`;
36
+ cursor += match[0].length;
37
+ } else if (open) {
38
+ result += `@{${open[1] ?? ""}${open[2]}${open[3] ?? ""}}`;
39
+ cursor += open[0].length;
40
+ } else {
41
+ result += source[cursor++];
42
+ }
43
+ }
44
+ return result;
45
+ }
46
+
47
+ /** Change Script raw bodies only; package references and generation prompt tags stay untouched. */
48
+ export function migrateSvml(source) {
49
+ let result = "";
50
+ let cursor = 0;
51
+ const token = /<!--[\s\S]*?-->|<(?:[A-Za-z_][\w.-]*:)?script(?=[\s/>])(?:"[^"]*"|'[^']*'|[^'">])*>/gu;
52
+ for (let match = token.exec(source); match; match = token.exec(source)) {
53
+ if (match[0].startsWith("<!--") || match[0].endsWith("/>")) continue;
54
+ const name = /^<([^\s>]+)/u.exec(match[0])[1];
55
+ const close = `</${name}>`;
56
+ const start = match.index + match[0].length;
57
+ let end = start;
58
+ while (end < source.length) {
59
+ if (source[end] === "\\") { end += 2; continue; }
60
+ if (source.startsWith("<!--", end)) {
61
+ const stop = source.indexOf("-->", end + 4);
62
+ if (stop < 0) throw new Error("Unclosed Script comment");
63
+ end = stop + 3;
64
+ continue;
65
+ }
66
+ if (source.startsWith(close, end)) break;
67
+ end++;
68
+ }
69
+ if (end >= source.length) throw new Error(`Missing ${close}`);
70
+ result += source.slice(cursor, start) + migrateScriptBody(source.slice(start, end)) + close;
71
+ cursor = end + close.length;
72
+ token.lastIndex = cursor;
73
+ }
74
+ return result + source.slice(cursor);
75
+ }
76
+
77
+ if (process.argv[1] && import.meta.url === pathToFileURL(resolve(process.argv[1])).href) {
78
+ const args = process.argv.slice(2);
79
+ const flags = new Set(args.filter(arg => arg.startsWith("--")));
80
+ const paths = args.filter(arg => !arg.startsWith("--"));
81
+ if (paths.length !== 1 || [...flags].some(flag => !["--write", "--body"].includes(flag))) {
82
+ console.error("Usage: node migrate-0.2.mjs <source.svml> [--body] [--write]\nDefault: print migrated SVML. --body: input is a raw Script body. --write: replace the named file.");
83
+ process.exitCode = 2;
84
+ } else {
85
+ const source = await readFile(paths[0], "utf8");
86
+ const migrated = flags.has("--body") ? migrateScriptBody(source) : migrateSvml(source);
87
+ if (flags.has("--write")) await writeFile(paths[0], migrated);
88
+ else process.stdout.write(migrated);
89
+ }
90
+ }