fix: normalize prompt_tokens to include cache tokens across Anthropic providers

[guoyangzhen]

Problem

When Portkey normalizes Anthropic model responses to OpenAI schema, prompt_tokens has inconsistent semantics depending on which provider is used:

Provider	prompt_tokens includes cache tokens?
anthropic (direct)	❌ No
vertex-ai (Anthropic)	❌ No
bedrock (Anthropic)	✅ Yes

The OpenAI convention (which Portkey normalizes to) is that prompt_tokens includes cached tokens, with the breakdown available in prompt_tokens_details.cached_tokens.

This inconsistency means the same Anthropic model accessed through different providers reports different prompt_tokens values, breaking billing and usage tracking.

Fixes #1564

Solution

Normalize prompt_tokens to always include cache_read_input_tokens + cache_creation_input_tokens, matching Bedrock's (correct) behavior and the OpenAI convention.

Changes

1. Anthropic direct (non-stream) — src/providers/anthropic/chatComplete.ts:613
   • prompt_tokens now includes cache creation and read tokens
2. Anthropic direct (streaming) — same file, message_start handler
   • Same fix applied to streaming chunk
3. Vertex AI Anthropic (non-stream) — src/providers/google-vertex-ai/chatComplete.ts:899
   • Same fix
4. Vertex AI Anthropic (streaming) — same file
   • Same fix

Before

prompt_tokens: input_tokens,  // excludes cache tokens

After

prompt_tokens: input_tokens + cache_creation_input_tokens + cache_read_input_tokens,

Testing

Verified the change matches Bedrock's existing behavior at src/providers/bedrock/chatComplete.ts:551-554:

prompt_tokens:
  response.usage.inputTokens +
  cacheReadInputTokens +
  cacheWriteInputTokens,