fix: refund users when model returns zero tokens with non-zero cost

[sh1ftred]

Problem

When a model produces empty output (zero tokens reported) but the cost calculation returns a non-zero USD cost, the previous behavior was to charge the full amount as output_msats. This means users were being billed for responses where the model produced nothing.

Fix

In the else branch of calculate_cost() (when no tokens are reported):

• Before: output_msats = cost_in_msats (user gets charged the full amount)
• After: Set input_msats = 0, output_msats = 0, cost_in_msats = 0 — issuing a full refund to the user
• Added a logger.warning with usd_cost and model details for debugging/monitoring

File Changed

• routstr/payment/cost_calculation.py

Testing

The change is a simple conditional branch update. When usage.input_tokens and usage.output_tokens are both 0 but usd_cost > 0, the user now receives a full refund instead of being charged.

[9qeklajc]

some upstream will have the problem of not setting the token count for some reasons even when the user gets a response which in that case implies that the request is free

[shroominic]

Heads up: I opened a clean re-implementation of this fix against current main — #561. This PR is 110 commits behind a refactored main (conflicting) and ships no tests; the new one re-applies the fix to the current _calculate_from_usd_cost code, tightens the gate so legitimately-billed cache-read-only turns aren't refunded (this PR refunds whenever prompt+completion==0, which over-fires on cache-only turns), and adds regression tests (truly-empty → refunded, verified fail-without-fix; cache-only → still billed). Suggest closing this in favour of #561 once reviewed.