Fix DeepSeek v4 Claude pricing through stale runtime cache

[ozymandiashh]

Summary

Addresses #365 by pricing DeepSeek v4 Claude-compatible sessions instead of reporting them as $0.

Root cause

CodeBurn relies on LiteLLM pricing data, but deepseek-v4-pro and deepseek-v4-flash are not yet available in the bundled/runtime LiteLLM data used by current CodeBurn. A fresh 24-hour runtime pricing cache can also replace the bundled snapshot with data that still lacks these models, so adding snapshot entries alone is not enough.

What changed

• Add temporary DeepSeek v4 manual pricing entries to the LiteLLM bundle until upstream LiteLLM includes them.
• Merge bundled snapshot fallback entries into cached or freshly fetched LiteLLM pricing only when the runtime data lacks a model.
• Add DeepSeek v4 display names for shared model reporting and the Claude provider.
• Add an end-to-end CLI regression fixture for Claude JSONL sessions with a stale runtime pricing cache.

Validation

I ran the same Claude JSONL fixture against origin/main and this branch. The fixture contains two Claude-compatible assistant calls using the reported DeepSeek v4 model ids, plus a fresh runtime LiteLLM cache that does not contain those models.

Input shape:

provider: claude
model: deepseek-v4-pro
input_tokens: 2,477,914
output_tokens: 762,994
cache_read_input_tokens: 258,556,928

model: deepseek-v4-flash
input_tokens: 1,552,573
output_tokens: 353,914
cache_read_input_tokens: 48,388,608

Before, on origin/main, CodeBurn parsed the calls and tokens but priced both models at zero:

deepseek-v4-pro:   calls 1, cache reads 258,556,928, cost 0
deepseek-v4-flash: calls 1, cache reads 48,388,608, cost 0
TOTAL:             calls 2, cost 0

After, on this branch, the same input produces non-zero costs and the expected display names:

DeepSeek v4 Pro:   calls 1, cache reads 258,556,928, cost 2.678966
DeepSeek v4 Flash: calls 1, cache reads 48,388,608, cost 0.451944
TOTAL:             calls 2, cost 3.13091

That before/after path is now covered by tests/cli-deepseek-v4-pricing.test.ts, which runs the CLI against the realistic Claude fixture and stale runtime cache.

Notes

• DeepSeek v4 Pro uses the currently documented 75% discounted price that DeepSeek lists as active until 2026-05-31 15:59 UTC.
• This overlaps with fix: add DeepSeek v4 Pro/Flash pricing and fix snapshot survival through runtime cache #366, but adds the stale-runtime-cache CLI proof and uses the current discounted Pro price.

[iamtoruk]

Hey @ozymandiashh I have just noticed that there was a similar PR but yours was better for next time, rather than having a duplicate PR if we can suggest changes on the existing PR, it would be better. :D