architecture

Parent: lucerna · Repo root: ~/Develop/projects/Otium/

Source of truth

Code is authoritative; ops/ARCHITECTURE.md (2026-05-23) has drifted on almost every infra claim — all confirmed against code:

• Desktop is Electron, not "Tauri 2" (desktop/package.json electron 33 + electron-builder + better-sqlite3; main/* import from 'electron'). A Tauri→Electron migration happened after the doc; many wa-sqlite/Tauri comments in desktop code are stale copy-paste.
• Desktop local store is native better-sqlite3 via IPC, not wa-sqlite (the doc's "wa-sqlite cache" is only true for web). The schema block is FULL_SCHEMA_SQL reapplied every boot; the PRAGMA user_version migration framework survives only on web.
• Bearer-token auth is shipped on desktop, not "planned" (setSessionProvider, Authorization: Bearer).
• Still aspirational: the library hybrid slice (quota/local-move), real per-user scoping (every desktop row uses the literal 'local' user + wipe-on-session-change), and the advertised browser-plugin/mobile shells (do not exist).
• landing/ is an empty stale clone; the real site is LandingPage/ (Payload CMS). Identity is dual/in-flight (Appwrite service + better-auth tables both present).

Stack (verified)

• Domain core: framework-agnostic TS (@lucerna/core, peer-dep only zustand, tsup).
• Shared UI: @lucerna/design-system (React 19, Radix, CVA, Tailwind 4).
• Shells: Next.js 15 (web) · Electron + Vite (desktop) · Next.js 15 + Payload CMS 3 (landing).
• Backends: Go 1.26 + echo (auth = Appwrite-backed; mainbackend = domain API + pgx + google/wire, owns the SRS) · Python + FastAPI + spaCy (nlp-api, DE+EN).
• Data: Postgres; client SQLite (wa-sqlite web / better-sqlite3 desktop); flexsearch. Cohesion via a private Verdaccio registry; Coolify deploy; R2 assets.

The reading pipeline (the spine)

Ingest is 100% client-side; the server stores the parsed result. The load-bearing idea is the content-keyed, two-coordinate sentence.

• Parse (core/src/upload/parseEpub|parsePdf|parseText): EPUB via JSZip → OPF/spine → HTML-stripped chapters joined; TOC from NCX/EPUB3-nav with cumulative char offsets; rejects scanned/thin books.
• contentHash (upload/contentHash.ts, MD5, deliberately legacy-compatible) is the join key — vocab/notes/sentence-cache all anchor to it, so re-upload relinks.
• CachedSentence (persisted unit, key (contentHash, sentenceIdx)) carries two coordinate systems — raw source offsets and display offsets the user reads — plus displayToRaw[]. This is the backbone: see raw-display-cursor.
• Page is a transient runtime aggregate (only the current page in RAM); layout decides which sentences share a pageNumber — see dynamic-pagination (DOM-measured, reflows on font/viewport).
• On top of the same sentence stream: full-text-search (per-book FlexSearch, disk-persisted) and, on desktop, tts-audiobook (local Piper audiobook with raw-offset-anchored karaoke). Ingestion that feeds all of this: upload-ingestion.

Reading domain model (plain terms)

Book (full text + ISBN provenance) → split into CachedSentences → read page by page (Page, transient) → position/marks stored content-anchored, not geometry-anchored so they survive font/layout changes: Anchor (current position = ≤200-char sentence prefix), Bookmark, ReadingNote (highlight with book-absolute raw offsets), VocabularyEntry (lemma + all inflected forms[] + POS/VerbTable/NounInfo + context + SRS state). POSCategory (10-way taxonomy, colors, 11-language labels, alias folding) and VerbTable (conjugation grid) are the linguistic value-objects.

Poly-repo shape

No monorepo tool — ~12 sibling repos cohere via the Verdaccio registry.

Key designs (the pillars)

• key-designs — the full index. Reading substance: dual-coordinate backbone, dynamic-pagination, all-forms-underline, vocabulary-loop, full-text-search, tts-audiobook. Linguistic core: linguistic-two-engines, upload-ingestion. Architecture pillars: cross-platform-seams, offline-first-outbox, backends-three-services.