chore(weave): add manual cost overrides and update costs 5/19

[jwlee64]

Description

Introduces a manual_costs.json file and a corresponding fetch_manual_costs() function to allow manually curated cost overrides for models that litellm and modelsBegin.json do not cover or cover incorrectly. Manual costs take precedence over both litellm and modelsBegin sources when merging in main().

New manual costs

• gemini-3.1-flash-lite (via Gemini, Vertex AI, and OpenRouter)

The fetch_manual_costs() function gracefully handles missing or malformed files by returning an empty dict and printing a warning, so the pipeline continues running from litellm and modelsBegin sources.

Ran make update_costs and make update_model_providers after adding the manual costs file

Testing

Unit tests were added for fetch_manual_costs() covering:

• Missing file returns empty dict without crashing
• Malformed JSON is tolerated and returns empty dict
• Well-formed entries load correctly while malformed entries are skipped
• When a model has multiple historical entries, the most recent one is used
• Manual costs correctly override litellm costs for the same model ID in the main() pipeline

Existing test_main tests were updated to mock fetch_manual_costs to maintain test isolation.

[codecov]

Codecov Report

❌ Patch coverage is 90.00000% with 3 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
weave/trace_server/costs/update_costs.py	90.00%	3 Missing ⚠️

📢 Thoughts on this report? Let us know!

[wandbot-3000]

Preview this PR with FeatureBee: https://beta.wandb.ai/?betaVersion=fd637308e206d7cc47808a14c5f53ed6e9d40497

[gtarpenning]

what happens if both litellm reports and we manually enter? who wins? probably should be litellm?

[jwlee64]

It was the manual costs, but makes sense to me to default to litellm