feat: auto-unload model after 30min inactivity + model picker on first launch

[Swallow74]

Summary

Two features for better model lifecycle management:

1. Auto-unload after inactivity

• Unloads the Gemma 4 model from GPU memory after 30 minutes of inactivity
• Automatically reloads the model when the user sends a message or reopens the chat
• Frees ~500MB-1.5GB of VRAM when not in use

2. Model picker on first launch

• Instead of immediately downloading the default model, the user sees a model picker
• Choose between E2B (~500MB) and E4B (~1.5GB)
• Selection is persisted in chrome.storage.local — only shown once
• Subsequent opens auto-load as before

Files changed

File	Change
offscreen/model-host.ts	Added lastModelId — persists across unload/reload
entrypoints/offscreen/main.ts	Added 30min inactivity timer + auto-reload in agent:run
content/chat-overlay.ts	Added model picker UI (CSS + DOM + show/hide)
entrypoints/content.ts	Check gemma_model_chosen flag, show picker or auto-load
shared/messages.ts	Added unloaded to ModelStatusMessage status union

[kessler]

hi @Swallow74
First, thanks for this PR
Quick question, those downloaded model sizes, where did you get them from ?

[Swallow74]

Thanks for checking!

I verified the sizes directly from the Hugging Face model repos. The app loads the q4f16 ONNX quantized format via Transformers.js, so I added up the actual onnx_data file sizes for that variant:

E2B (~3.4GB):

• decoder_model_merged_q4f16.onnx_data → 1.52 GB
• embed_tokens_q4f16.onnx_data → 1.59 GB
• audio_encoder_q4f16.onnx_data → 171 MB
• vision_encoder_q4f16.onnx_data → 99.2 MB
• Total: ~3.38 GB → rounded to ~3.4 GB

E4B (~5.2GB):

• decoder_model_merged_q4f16.onnx_data → 2.07 GB + 812 MB
• embed_tokens_q4f16.onnx_data → 2.02 GB
• audio_encoder_q4f16.onnx_data → 172 MB
• vision_encoder_q4f16.onnx_data → ~100 MB
• Total: ~5.17 GB → rounded to ~5.2 GB

I originally had the default ONNX sizes (~500MB and ~1.5GB), which seemed wrong the first time I actually downloaded a model, so I updated them to match the q4f16 files that Transformers.js pulls down.

Feel free to double-check at:

• https://huggingface.co/onnx-community/gemma-4-E2B-it-ONNX/tree/main/onnx
• https://huggingface.co/onnx-community/gemma-4-E4B-it-ONNX/tree/main/onnx