@graphlearning/shell 0.7.0 → 0.10.0

package/README.md

@@ -88,6 +88,23 @@ The recorder (`scripts/record-course.mjs` in each content repo) drives these, so
 - `?capture=1` — suppress every interactive control, leaving a clean frame
 - `window.__scene.plan()` — the active course's sections, for the recorder to walk
 
+## Keyboard
+
+Interactive only — `?capture=1` suppresses nothing here because none of it fires without a reader.
+
+| Key | What it does |
+| --- | --- |
+| `←` / `→` | seek the narration −10s / +10s **when the section has a clip** (exactly when the scrub bar is on screen); otherwise page to the previous/next section |
+| `Shift`+`←` / `→` | always page, clip or no clip |
+| `Space` | play/pause the narration |
+| `Esc` | back to the catalog (and from the catalog, up to graphl.in) |
+| `←` / `→` / `Home` / `End` | with the scrub bar focused: seek ∓10s / to the start / to the end |
+
+The arrows seek rather than page because the footer's `‹` `›` buttons are the pager's real home and
+`Esc` still leaves — while the only thing a *listening* reader has no control for is a 10s jump.
+Where there is no clip to seek (the catalog, a bare scene view, a repo whose wavs are not generated
+yet) the old paging meaning stands, so the keyboard never goes dead.
+
 ## Develop
 
 `npm run dev` serves a fixture harness on :5176 — two throwaway courses with a diagram scene, a code
@@ -113,12 +130,27 @@ They are `bin` entries, wired through each repo's `npm run`:
 | command | what it does |
 |---|---|
 | `npm run record` | one course → a 3840×2160 MP4 |
+| `npm run record:all` | every course in syllabus order, waiting for narration to land |
 | `npm run record:reels` | one course → portrait reels |
 | `npm run shots:4k` | one PNG per section, no ffmpeg |
 | `npm run thumb` | a branded YouTube thumbnail |
 | `npm run gen:desc` | video descriptions + chapters |
 | `npm run gen:audio` | the narration manifest |
 
+`record:all` is the only one whose repo script is not just the bin: it needs a sleep wrapper, because
+a full concept is a multi-hour batch and the machine going idle mid-run kills it.
+
+```json
+"record:all": "caffeinate -ims graphl-record-all"
+```
+
+It exists because narration arrives DURING a batch — the Colab + Chatterbox pass commits wavs one
+section at a time — so before each course it pulls the repo's branch and counts that course's wavs
+against `scripts/audio-manifest.json`, waiting (`--wait`, default 240 min) for the missing ones.
+A course is recorded only when it is complete; an incomplete one is skipped and named in the summary
+rather than half-recorded into `record`'s 3s-silence fallback. Re-running is cheap — finished courses
+are segment-cache hits.
+
 **Capture is not realtime.** `record` and `record:reels` screencast one short window per section and
 loop it over the narration (`-stream_loop`), because the only moving thing in a frame is the engine's
 edge pulse and that pulse has a 2.4s period — and a pulse crosses its whole edge in exactly one

package/dist/NarrationBar.d.ts

@@ -1,3 +1,4 @@
+export declare const SEEK_S = 10;
 export declare function NarrationBar({ audio }: {
     audio: HTMLAudioElement | null;
 }): import("react").JSX.Element | null;