feat: local image generation

[MegalithOfficial]

Local Image Generation: stable-diffusion.cpp engine, Playground, LoRA ecosystem, and guided tours

This PR ships fully local, on-device image generation end to end: a managed stable-diffusion.cpp runtime, a curated hardware-aware model catalog, a dedicated Playground for freeform generation, a LoRA library with an integrated CivitAI browser, deep integration into the existing avatar/scene flows, guided tours across every new surface, and full localization. Local generation is desktop-only by design; mobile keeps the existing provider-based flows and never touches sd.cpp.

Stable Diffusion Engine

• Managed sidecar: stable-diffusion.cpp runs as a managed sd-server sidecar with a native JSON API, forwarded logs, and lifecycle handling.
• Versioned engine manager: install, switch between, and remove multiple engine builds. The build catalog is resolved at runtime; the best variant for the detected GPU (Vulkan/CUDA/CPU) is preselected automatically, with the detection result shown inline.
• GPU-aware runtime manager: hardware is probed once and reused for build selection, model fit checks, and compute policy decisions.
• Upscaler: one-tap install of the RealESRGAN model that powers the Upscale action and the high-resolution pass.

Local Model Catalog

• Curated, hardware-checked model families: Z-Image (Turbo + base), FLUX.2 Klein (4B, 9B, 9B Base), Krea 2 (Turbo + RAW), and Qwen Image Edit 2511, each with per-variant quantizations, licenses, and sensible generation defaults.
• Verified bundle installer: Select & Download resolves the full component set (diffusion GGUF, text encoders, VAE) from Hugging Face, with stall guards instead of brittle total-request timeouts, and per-role type badges on installed files.
• Custom models: local image models are keyed by diffusion GGUF path; component files can be hand-picked from disk or from already-downloaded files, with auto-detection in the model editor.
• Runtime feature work: generation progress events, cancellation, inpainting, upscaling, step caching, expanded sampler/scheduler/CFG/size controlompatibility checks.
• The HF token moved to Local Runtime Defaults with guidance for gated model access.

Playground

A new desktop-only /playground route reachable from Settings → Image Generation:

• Three-pane layout: prompt pane (prompt, negative prompt, img2img init image), fullscreen feed carousel in the center, and a settings pane (meps, CFG, sampler, scheduler, seed, hires fix, batch, LoRA stack). Side panes are drag-resizable; on narrow windows the panes become bottom sheets.
• Write-ahead generation history: every generation is persisted with its full parameters and materialized seed, so history survives crashes and restarts.
• Feed carousel: deterministic wheel paging, snap scrolling, measured skeletons at the target aspect ratio (including batch grids), and lazy older-page loading.
• Remix actions on every card: copy prompt, reuse seed, regenerate, upscale, send to img2img, expandable prompt, per-image overlay actions for batches, and delete with optional file cleanup.
• Init image sources: pick from disk or from any previous generation.

LoRA Library + CivitAI

• New LoRA Library page with two tabs: Library (installed files with size, architecture badge, trigger keywords, keyword editing, CivitAI keyword discovery, delete) and Browse CivitAI (search, sort/period, base-model filter chips, NSFW-safe Pure-mode filtering, version/file pickers, queued downloads that land
  directly in the library).
• CivitAI backend: search, model detail, auth (token save/validate/clear stored in the local DB), and download queueing through the shared download pipeline.
• Only officially supported models are surfaced: search results, filter chips, and detail versions are all gated by the engine catalog itself (derived from the profile → LoRA architecture mapping, so new catalog families become browsable automatically). SD 1.x/2/XL, Pony, Illustrious, NoobAI, and Flux.1 content no longer appears.
• Full result pages: the search command now pages CivitAI (up to 5 fetches of 100) until the requested count of supported results is met, so the first page arrives full instead of shrinking to a handful after filtering.
• Missing-token warning: the browse tab checks token status and shows an inline warning (missing or expired) with an Add/Replace action; public downloads still work without one.
• Always-active LoRAs: LoRAs without trigger keywords are supported and labeled as always active.
• Flux LoRA application fixes and architecture detection from safetensors metadata with a compatibility cache.

Chat / Avatar / Scene Integration

• Scene prompts are tailored for local image models, with split image protocols per model type.
• Avatar generation supports local models (AI-editing disabled where the model can't do it).
• Deleted local image models no longer re-register themselves on listing.
• stable-diffusion.cpp is branded and kept selectable as a provider.

Settings & Discoverability UX

• Image Generation page: Playground and Stable Diffusion Engine now share a compact side-by-side tools row at the top (desktop only), so Generation/Prompting config is no longer pushed below the fold.
• Models page: the Model Browser / Local Runtime Defaults / LoRA Library shortcut row switched from dashed "empty-state" styling to the app's solid nav-card idiom with one-line descriptions.
• LoRA Library: tabs, installed count, and Import file merged into a single header row.

Guided Tours

• Five new tours on the existing GuidedTour engine: Image Generation settings, Playground, LoRA Library, CivitAI Browse (fires on first visit to the Browse tab), and Stable Diffusion Engine, plus a new CivitAI-token step in the existing Runtime Defaults tour.
• Engine upgrades: per-step desktopOnly flag (steps are filtered on mobile instead of stalling on missing targets) and optional flag (condip in ~400ms instead of 4s).
• Demo card: if the Playground feed is empty when the tour reaches the feed step, a realistic example card (image, prompt, model/size/seed chips, "Example" badge) renders in place of the empty state. The demo asset is dynamically imported only when that step is active and disappears with the tour.
• All tours replayable via window.__debug.resetTour(id) / resetAllTours().
• Demo card: if the Playground feed is empty when the tour reaches the feed step, a realistic example card (image, prompt, model/size/seed chips, "Example" badge) renders in place of the empty state. The demo asset is dynamically imported only when that step is active and disappears with the tour.
• All tours replayable via window.__debug.resetTour(id) / resetAllTours().

Desktop-only Gating

• The Playground card and SD Engine section are hidden on mobile; /playground and the LoRA Library render a desktop-only fallback if reached directly; sd.cpp commands are never invoked from mobile code paths.

i18n

• All new UI strings (tours, token warnings, playground labels, LoRA library, engine manager) added to en.ts and translated across all 21 locales in each file's established register and terminology; zh-Hans/zh-Hant written independently. Brand terms (CivitAI, LoRA, Playground, Stable Diffusion, ESRGAN) untranslated.

Verification

• bun run check (tsc + cargo check) green throughout.
• Tours exercised via the debug reset helpers on populated and empty states; missing-token banner verified against saved/missing/expired token states; mobile gating verified via the platform fallback paths.