llm_anthropic: emit per-call token usage into flow trace

[joshuadarron]

Problem

apaevt_flow events from llm_anthropic invocations include duration but not token counts. This makes it impossible to distinguish "this role takes longer because it processes more tokens" from "this role takes longer per token" without re-instrumenting from scratch.

In a coding-agent tracer run, eng2 consumed 60 LLM calls at 12.4 s avg vs eng1's 31 calls at 3.9 s avg. Without usage data we can't tell whether eng2 is context-bloated, output-heavy, or model-slow on its specific prompt shape.

Proposed fix

After each ChatAnthropic call returns, extract response.usage_metadata (or LangChain's AIMessage.usage_metadata) and attach to the trace.result payload of the apaevt_flow op:leave event:

{
  "input_tokens": 1234,
  "output_tokens": 567,
  "cache_creation_input_tokens": 0,
  "cache_read_input_tokens": 0
}

Free piggyback on the same flow event — no new event type, no new schema, no extra round-trip.

Acceptance

• apaevt_flow op:leave for any llm_anthropic invoke includes a usage dict.
• Caching verification (companion issue on prompt caching) becomes inspectable from the tracer file alone.
• Optional: also record on llm_openai, llm_bedrock, etc. if upstream usage shapes are similar — out of scope for this issue, separate ticket.

Suggested labels

enhancement, observability, nodes/llm_anthropic

[github-actions]

[2001anshkaushik]

I've submitted a PR for this issue that fulfills the flat-key acceptance criteria: #867

I noticed the alternative PR (#799) utilizes instance state mutation (self._last_usage inside ChatBase) to pass the telemetry. In high-concurrency environments, storing request-specific metrics on the instance level introduces a high risk of race conditions and data bleed across concurrent async requests.

To avoid this, my implementation takes a strictly stateless approach. By introducing a ChatResponse dataclass and using an Adapter Pattern inside _chat_with_retries, the metadata is tunneled entirely within the function call scope. This ensures absolute thread safety and natively protects downstream legacy nodes (Gemini, Watson) from structural breakages without requiring complex normalization helpers.