Add the Next.js web app and shared lesson content.

Move lessons out of the C sources into JSON so the web client can use the same material as the desktop app, and point release builds at desktop/.

Co-authored-by: Cursor <cursoragent@cursor.com>

content/README.md

@@ -0,0 +1,80 @@
+# content/
+
+Shared lesson content for the whole project. One folder, one file per lesson —
+instead of lesson text being buried inside each unit's C source file.
+
+`schema.ts` is the single description of every shape here. Both the extractor in
+`tools/` and the web app in `web/` type against it, so a change there shows up as
+a type error on both sides.
+
+## Layout
+
+```
+content/
+  schema.ts              TypeScript types for everything below
+  themes.json            10 palettes, dark + light (24 colors each)
+  subjects.json          the 13 subjects, their i18n keys and routes
+  paths.json             serpentine unit paths per subject
+  manifest.json          subject and lesson ids
+  i18n/
+    ui.json              UI chrome strings, keyed by identifier
+    content.json         lesson strings, keyed by the Czech text
+  lessons/
+    deutsch/unit-01.json … unit-03.json
+    networks/unit-01.json … unit-27.json
+    hardware/unit-01.json … unit-03.json
+    czech/mluvnice.json, orwell-1984.json, fuks-spalovac-mrtvol.json
+```
+
+`paths.json` holds one entry per path. A subject usually has exactly one, but
+Czech has two: the top-level area picker and the `reading` branch listing the
+books, distinguished by the `branch` field.
+
+## Lesson shape
+
+A lesson holds theory `slides` and a list of `tasks`. Each task is one bubble on
+the unit map and contains `items` of a single `kind`:
+
+| kind | what it is |
+| ---- | ---------- |
+| `choice` | multiple choice, `correct` indexes `options` |
+| `fill` | sentence split into `segments` with a gap between each; `input` picks dropdown or text |
+| `dialogue` | dialogue lines with one missing word each |
+| `assembly` | build a sentence from scrambled words |
+| `assign` | sort labelled cards into named groups |
+| `free` | open question with a model answer |
+| `typed` | type the answer, `answers` lists accepted spellings |
+| `fields` | one prompt, several labelled fields (Czech grammatical categories) |
+| `reveal` | self-assessed: write an answer, then compare it with the model |
+| `hangman` | guess the word letter by letter |
+| `profile` | "Steckbrief" line with a model answer |
+| `table` | editable grid, some cells prefilled |
+| `subnet` | VLSM subnetting worksheet |
+
+`fill` absorbs every gap variant in the desktop app. A few flags steer it:
+`letters` marks gaps holding missing letters inside a word rather than whole
+words, `gapPool` picks the task's `pool` or `altPool` per gap (the "Ordne zu"
+question-word and answer columns), and `heading` starts a labelled group of rows.
+
+Two conventions carried over from the desktop app:
+
+- **Pipe-separated answers.** `"bin|heiße"` means either spelling is accepted.
+- **Lenient comparison.** Answers are matched after lowercasing, expanding
+  `ä ö ü ß` to `ae oe ue ss`, dropping accents and ignoring punctuation, so Czech
+  diacritics and German umlauts are optional. See `web/src/lib/answers.ts`.
+
+Titles come either as a literal `title` or as a `titleKey` to look up in the
+i18n tables, because German unit names are identical in both languages while
+network lesson names are translated.
+
+## Regenerating
+
+The C sources in `desktop/` are still where content is authored, so this folder
+is a generated one-way export:
+
+```bash
+npm run content     # from the repo root, or: node tools/build-content.mjs
+```
+
+The extractor parses the C initializers directly and throws on any shape it does
+not recognise, so a bad export fails loudly rather than emitting empty lessons.