Cap max_output tokens to 15%-20% of input tokens

[isidorn]

As proposed by Logan: "if we think 64k is too much we should come up with some heuristics as to what to reserve for output. Maybe output should be at max 20% input?"

15%-20% are good numbers.
For example, sonnet 4 has 200k, 16k = 12.5%
Gemini 2.5pro has 1M input, 64K output = 15%

Thus my recommendation is that we cap max_output at 15% of input.

fyi @roblourens

[roblourens]

Seems like this isn't working because we still allow 64k output tokens, and it's hard to debug because gemini doesn't show up while debugging but I'll look into it

[roblourens]

I see that we change the number of tokens that we reserve, but do we also need to override capabilities.limits.max_output_tokens or does that not matter? We log that max output tokens is 64k but maybe that doesn't matter and the only thing that matters is adjusting max_prompt_tokens? I'll mark it verified for now but you can double check when you're back. @lramos15