feat: add LM Studio backend, draggable + per-session-hideable gem icon

[guberm]

Summary

Adds two user-facing features and a pnpm build fix:

1. Local model via LM Studio (over a link) — run inference against a local OpenAI-compatible endpoint instead of in-browser WebGPU.
2. Movable gem icon — drag the floating icon anywhere; position persists.
3. Hide the gem icon for the current session on the current site — a temporary, per-hostname hide that clears on browser restart (distinct from the permanent "Disable on this site").

1. LM Studio / remote backend

A third model option, LM Studio (local), appears in the gear-icon settings. When selected, an endpoint panel lets you set the base URL (default http://localhost:1234/v1), an optional model name, optional API key, and a context limit.

• The agent already builds the full Gemma-formatted prompt, so the remote backend POSTs it raw to {baseUrl}/completions with stream: true and parses the SSE deltas — the same special-token tool/thinking protocol applies as the WebGPU path. A Gemma model loaded in LM Studio is recommended for tool use.
• Special-token parsing (thinking blocks, tool calls, visible text) was extracted from model-host.ts into a shared GemmaStreamFilter so both backends behave identically.
• load() pings /models so you get immediate feedback if LM Studio's server isn't running.
• Settings persist across sessions; switching back to a WebGPU Gemma model is one dropdown change.
• Limitation: image/audio tools (screenshots) work only on the WebGPU backend, not over the LM Studio link.

New: offscreen/remote-model-host.ts, offscreen/stream-filter.ts. The offscreen agent loop now swaps between an in-browser host and the remote host via an activeHost reference; the background reads/persists the endpoint config and forwards it on model:load / model:switch.

2. Draggable icon

Press-and-drag the gem icon (pointer events, 4px threshold to distinguish a drag from a click, clamped to the viewport). Position is saved to storage.local and restored on every page.

3. Per-session hide

A Hide gem icon (this session) button hides the icon per-hostname using chrome.storage.session (survives reloads, clears on browser restart). While hidden, reopen the chat with the toggle shortcut (Alt+G), where the button flips to Show gem icon. Required granting content scripts session-storage access in the background.

Build fix (pnpm)

wxt.config.ts resolved onnxruntime-web via require.resolve('onnxruntime-web'), which only worked under npm's flat node_modules. Now it resolves relative to @huggingface/transformers (its parent), so the build works under pnpm's nested layout.

Files

• shared/models.ts — lm-studio model + RemoteEndpointConfig type, isRemoteModel(), storage keys/defaults
• shared/messages.ts — remote:config message; remoteConfig on model load/switch
• offscreen/remote-model-host.ts, offscreen/stream-filter.ts — new
• offscreen/model-host.ts, entrypoints/offscreen/main.ts — backend swapping + shared stream filter
• background/message-router.ts, entrypoints/background.ts — config persistence + session-storage access
• content/chat-overlay.ts, content/gem-icon.ts, entrypoints/content.ts — settings UI, drag, session-hide
• wxt.config.ts — pnpm-compatible ORT resolve
• README.md — docs for all three features

Testing

• wxt build --mode development ✔ — extension builds into .output/chrome-mv3-dev.
• Manual: load unpacked, start LM Studio's local server with a Gemma model, select LM Studio (local); drag the icon; hide/show it for the session.

Notes for reviewers

• public/ort/ort-wasm-simd-threaded.asyncify.mjs changed because the pinned onnxruntime-web dev version copies a matching loader. Revert if you want to keep the previously committed loader.
• pnpm compile still reports pre-existing errors (chrome, navigator.gpu, transformers strict-null) because the repo doesn't declare @types/chrome / @webgpu/types. These predate this PR and don't affect the build. Happy to add those devDependencies in a follow-up if desired.