Fix LiteLLM cost model resolution

[bearsyankees]

No description provided.

[greptile-apps]

Greptile Summary

This PR adds provider-qualified LiteLLM model resolution and revises usage accounting to prefer observed costs over local estimates without counting both.

• Adds cached model-name normalization and pricing-table resolution.
• Uses resolved model names for callback and SDK usage estimates.
• Separates observed and estimated cost accumulators and preserves cost across hydration.
• Adds tests for model resolution, routed-provider estimates, observed-cost precedence, hydration, and zero-cost runs.

Confidence Score: 5/5

The PR appears safe to merge, with no concrete changed-code failure identified.

The new resolver preserves provider-qualified entries, declines ambiguous suffix matches, and the revised ledger avoids adding observed and estimated representations of the same calls while retaining persisted historical cost.

Important Files Changed

Filename	Overview
strix/report/pricing.py	Adds defensive, cached resolution from configured model identifiers to unambiguous provider-qualified LiteLLM pricing keys.
strix/report/state.py	Resolves callback model candidates before asking LiteLLM to estimate otherwise unavailable response costs.
strix/report/usage.py	Separates observed and estimated totals, uses resolved pricing names, and restores persisted aggregate cost for resumed runs.
tests/test_cost_tracking.py	Updates the callback fallback expectation to use the provider-qualified LiteLLM model name.
tests/test_pricing.py	Covers representative resolution, ambiguity, observed-cost precedence, hydration, routed estimates, and zero-cost behavior.

_{Reviews (1): Last reviewed commit: "Fix LiteLLM cost model resolution" | Re-trigger Greptile}