fix(instrumentation): all 4 usage sources are 'if' not 'elif'

[maltsev-dev]

Summary

extract_usage_from_response in src/nullrun/instrumentation/langgraph.py
walked the 4 token-extraction source branches as an elif chain:

if hasattr(response, 'usage_metadata'):
    # read from usage_metadata
elif hasattr(response, 'generations'):
    # read from generations[0][0].message
elif hasattr(response, 'usage'):
    # read from response.usage
elif hasattr(response, 'response_metadata'):
    # read from response_metadata.token_usage

A LangChain AIMessage can carry token info on multiple attributes at
once. When the first branch's hasattr returned True but the value
was an empty dict or 0/0/0 (streaming init state, some provider
wrappers), every subsequent elif was skipped -- including the one
carrying the real token counts. The SDK shipped tokens=0 to
/api/v1/track, making the LLM call invisible on the dashboard.

Reproducer (now pinned by the new test)

response.usage_metadata = {input_tokens: 0, output_tokens: 0, total_tokens: 0}
response.response_metadata = {token_usage: {prompt_tokens: 26, completion_tokens: 48, total_tokens: 74}}

# Pre-fix: ships tokens=0 to /track. LLM call invisible on dashboard.
# Post-fix: ships tokens=74, input_tokens=26, output_tokens=48.

Fix

Switched all 4 source branches from elif to plain if so each
one attempts its read; later branches naturally overwrite the zero
default when the earlier branch's value is empty. In practice
LangChain providers never put conflicting token counts on two
attributes of the same response, so "last-wins" is safe.

Tests

• New test_extract_usage_metadata_zero_response_metadata_real in
  tests/test_langgraph_callback.py pins the regression.
• All 39 test_langgraph_callback.py tests pass.
• All 68 test_extractors.py + test_instrumentation_phase41.py
  tests pass (no regression in the wire-shape normalization
  testers).
• 107/107 tests green in the touched files.

Coverage scope

The same extract_usage_from_response is the single source of
truth for all 5 framework integrations -- autogen, crewai,
llama_index, the LangChain callback path, and the OpenAI Agents
SDK tracer all import or re-use the same NullRunCallback /
extraction logic. Fix in this one place closes the bug for every
framework at once.

The 5 direct HTTP extractors in auto.py (_openai_extractor,
_anthropic_extractor, _gemini_extractor, _cohere_extractor,
_bedrock_extractor) parse the raw HTTP response body directly
with no chain -- they do not exhibit this bug pattern. Audited
separately.

Version

• 0.13.3 -> 0.13.4 (changelog entry appended in
  src/nullrun/__version__.py).
• Wire format unchanged.
• SDK_MIN_VERSION_FOR_V3 unchanged ("0.12.0"); backends on
  1.0.0 keep working unchanged.

Co-Authored-By: Hermes noreply@nous.research

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!