fix(ai-gateway): stop double-charging cache tokens in spend

domain.Usage.PromptTokens is cache-inclusive, but usageFromDomain mapped
it straight onto the spend wire's input_tokens while also sending the
cache read and cache creation buckets separately, so the rating side
charged cached tokens at both the input rate and the cache rate. Subtract
the cache buckets from input_tokens, matching the span emitter's
exclusive split.

Fixes langwatch#7018.

Author: feiiiiii5

services/aigateway/adapters/spendemitter/contract_test.go

@@ -110,7 +110,7 @@ func TestContractConfirmedPayload(t *testing.T) {
 	require.NoError(t, json.Unmarshal(records[0].Payload, &payload))
 	usage, ok := payload["usage"].(map[string]any)
 	require.True(t, ok)
-	assert.EqualValues(t, 869, usage["input_tokens"])
+	assert.EqualValues(t, 853, usage["input_tokens"])
 	assert.EqualValues(t, 207, usage["output_tokens"])
 	assert.EqualValues(t, 11, usage["cache_read_input_tokens"])
 	assert.EqualValues(t, 5, usage["cache_creation_input_tokens"])

services/aigateway/adapters/spendemitter/record.go

@@ -108,7 +108,7 @@ type ErrorPayload struct {
 // reasoning-token count today.
 func usageFromDomain(u domain.Usage, reasoning int) UsagePayload {
 	return UsagePayload{
-		InputTokens:         u.PromptTokens,
+		InputTokens:         u.PromptTokens - u.CacheReadTokens - u.CacheCreationTokens,
 		OutputTokens:        u.CompletionTokens,
 		CacheReadTokens:     u.CacheReadTokens,
 		CacheCreationTokens: u.CacheCreationTokens,