Was it a different level design

Yep. WebGL projects are effectively single-threaded without SIMD. Anything up to API contact is quite expensive (2 - 2.5x slower than desktop, before considering SIMD). A lot of the means of optimizing rendering come with a notable CPU overhead, and where cutting into our frame-time. Compounding this is the fact that URP is a simple forward renderer, so you pretty much incur the cost of geometry for each light in its vicinity. While URP allegedly has directional lightmaps, in practice there is no specular contribution from baked indirect lighting, requiring a number of dynamic fill lights for normal maps to have any contribution.

So, entirely new assets as to allow ~95% of a frame to be drawn with 3 materials, and a layout with shorter sight-lines. Not as pretty, but passable performance. TLDR; I wouldn't recommend WebGL for anything more than the simplest of games. We're pretty much calling this a learning experience, and getting it out of the door before moving onto a more serious desktop project.

maybe you could have one button throw the item and the other to drop it.

That's actually in already. Right clicking while carrying a book throws it, and will distract the enemy if it isn't in a pursuit state. Crouching/sprinting, the effect that has on sound propagation, lights effect on visibility etc; That all needs to be covered in a brief tutorial. We'll be looking into stunning the AI, though.

putting it lower should be all you need

For the next test, we'll see how people fair with a short textual tutorial. If that's not enough, then this, definitely.

do you mean signs within the actual environment

Yeah, colored indicators for the various sections, coupled with illuminated exit signs. A proper map will be included as well, along with local lighting in the start room to make it and the objectives board more noticeable.

Maybe define a limited array of colours in the shader and only generate and store an index value

That's certainly possible for the interactable books, not so much for the static ones though since GPU instancing is significantly more expensive than static batching in WebGL. Unfortunately, the overlay approach would reduce visual quality since you're effectively limited to shifting values, not hue. One could look into multi-channel masks, or curve based albedo reconstruction, but that's pretty time consuming to get right and possibly outside of our performance budget.

We actually have some book variations (both in shape and color), and there is still space in that texture atlas, so we could offer at least 8 design variations. We just forgot to randomize the mesh selection for objective spawns... Would you be happy with that?