Major Engine Update

This update has only modest visible changes (fixed some bugs, some UI Improvements) but has a variety of key engine improvements. These include:

• Retooling of the embedding engine. This is used for everything from selecting images to managing game state items like stamina and orgone, offloading dozens of tasks that need to be completed each turn from the computationally expensive LLM. Previously, the embedding model was used in the standard way: mean pool all token-level embeddings to generate a final embedding, then compare embeddings with dot-products to find similarities, such as images whose tags are more similar to the narration or to rank queries such as "<NPC> is interacting with you" to determine the current focus NPC.  Now, the individual token embeddings, as well as the latent model attentions, are used for lightweight (~30 ms per turn) but highly effective post-processing, identifying which parts of the text are relevant to each NPC, as well as the player, dramatically improving state management and making it possible to identify which NPC is doing what in a way that just wasn't possible with standard methods. You should see much better assignment of focus NPCs and NPC travel, management of stamina, etc. 
• In this process, I moved to a quantized embedding model, but found that the onnx webgpu implementation is still quite buggy, so I've moved to embedding model to wasm (CPU) only. It is quite fast with the quantized weights, though, so there shouldn't be a large performance hit even if you had a dedicated GPU.
• With these improvements in hand, I've rescaled stamina and orgone assignment. Recent versions of the game were extremely generous in both due to the difficulty in deciding when to add orgone or decrement stamina--I decided it was better to have a system that was generously incorrect than unfairly incorrect. Now both are managed more effectively, so you should see a bit more challenge, with consistent stamina drops due to climax and lower orgone generation outside of intense encounters. 
• I've greatly improved the performance of the app overall. In addition to API latency (out of my control), every turn previously took about 10+ seconds to process, encompassing the local embedding model, image loading, saving and loading, game logic, etc. It's down to about 2-3 seconds now, so you shouldn't have to watch a locked screen nearly so long anymore.
• Image history management is (hopefully) really fixed now.
• For openrouter LLMs, I've added automatic handling for models that don't handle structured requests, allowing you to use  many popular roleplay models that don't support these types of API calls, notably those from deepseek.