A full rework of the beta series, rebuilt around one goal: reading manga should feel seamless.

The full pipeline - detection, script gate, OCR, translation, inpainting, rendering - runs on-device in WebGPU Mode, with Gemini cloud and self-hosted API Mode as opt-in alternatives. Zero telemetry, zero accounts, zero proxy in-between.

🖱 Interaction: Pill, Hub & Auto-Translate
1. Floating Translate pill. Hover any page image (≥ 260px) and a pill appears at its top-left - right where your eyes already are. No context menus, no toolbar hunting, no leaving the page. After translation it stays put as an action hub: fix a line, nudge a box, compare with the original, export the page.
2. Skip bbox refining jumps straight from detection to translation for clean pages; Auto-translate turns whole chapters hands-free, rendering each page as it scrolls into view, one at a time so nothing stutters.
3. A strict status registry dedupes repeat triggers across hover, auto-translate, and popup actions - the same page never translates twice.

🔒 Privacy & Security Core
1. Zero telemetry, permanently. All telemetry pathways, data-sharing toggles, and reporting endpoints are deleted - not disabled. WebGPU Mode transmits zero bytes anywhere.
2. SSRF-gated image proxy. The background image proxy allowlists http(s) only and rejects credentials, loopback, private/link-local IPs, and internal schemes.
3. Strict translation schemas. Every backend (WebLLM, server, Gemini) validates translations[] type and exact bbox-count match before anything is painted.
4. SHA-256 model verification. Downloads are checksummed against a known-hash manifest; mismatches evict the cache entry and abort.

🔍 Detection & OCR
1. RT-DETR default. Comic Bubble Detector (bubbles + free text) is now the default; Comic Text Detector with pixel-mask seeding stays selectable. YOLO26 models were removed - old stored selections auto-migrate to the new default.
2. 11 PaddleOCR language packs. Latin + Chinese/Japanese ship during onboarding; the rest download on first encounter while the overlay names the model. The script gate auto-switches packs per page under Auto-Detect.
3. Three OCR engines. PaddleOCR multilingual (default), PP-OCRv6 Manga (lightweight Japanese fine-tune), Manga-OCR (Japanese flagship for vertical text).
4. Script gate. A 3.7 MB on-device model holds back sound effects and wrong-language lettering; every hold-back keeps its recognized text and offers one-click Translate anyway.

🎨 Inpainting: Fast & Quality
1. Fast ladder (default, model-free). Fitted text-shaped masks through planar fill → bilateral denoise → Telea, lightest passing rung wins, decline-gated with rollback.
2. Quality mode (opt-in). LaMa neural redraw first per region (~207 MB one-time download), automatic Fast fallback wherever LaMa declines - a strict superset of Fast.
3. Declined regions are left pixel-identical and flagged, never smeared.