Cost is 0 for deepseek-v4-* models with claude

[NihalJain]

[ozymandiashh]

I looked at #366. It is directionally right: it adds fallback pricing for deepseek-v4-* and also handles the runtime LiteLLM cache path, which is the important root cause here.

One thing to fix before merging that PR: the current DeepSeek pricing page lists deepseek-v4-pro at the active 75% discounted rate until 2026-05-31 15:59 UTC, so the current values in #366 overprice Pro by 4x:

#366 deepseek-v4-pro:
input  1.74e-6
output 3.48e-6
cache  1.45e-8

current documented discounted price:
input  4.35e-7
output 8.7e-7
cache  3.625e-9

I also opened #367 with the discounted Pro pricing and an end-to-end CLI regression for the reported behavior: the same Claude JSONL fixture with a fresh runtime LiteLLM cache missing these models reports cost 0 on origin/main, then reports non-zero costs on the branch.

So either #366 should update the Pro rates and add a CLI-level stale-cache regression, or #367 can be used as the follow-up fix.