fix: 0.8.3 — close silent zero-billing across langgraph + init-ordering

[maltsev-dev]

Three coordinated defenses against the same class of bug 0.8.2 closed on the httpx path: llm_call events reaching the backend without a model field were silently recorded as ≈$0 (backend unwrap_or('default') + DEFAULT_RATE). 0.8.3 closes the langgraph callback path and the init-ordering hazard, and promotes the missing-model wire failure from WARN to fail-LOUD.

1. langgraph callback path (src/nullrun/instrumentation/langgraph.py)
_extract_model_from_response now consults response.llm_output FIRST — that's where langchain-openai 1.x puts the date-suffixed model id (gpt-4.1-mini-2025-04-14). The previous chain led with response_metadata, which langchain 1.x leaves empty on the AIMessage inside generations[0][0].message. Without this promotion every OpenAI-via-LangChain 1.x call silently zero-billed. Also adds a "any key containing model" sweep inside llm_output for non-OpenAI wrappers (proxies, custom chat models).

2. Init-ordering hazard (src/nullrun/instrumentation/auto.py)
patch_httpx's class-level __init__ wrap only catches Clients created AFTER it is installed. Users that build ChatOpenAI(...) before nullrun.init(api_key=...) get a pre-existing httpx.Client that the patch never sees. We now sweep gc.get_objects() once at patch install and wrap any pre-existing Client/AsyncClient whose transport isn't already a NullRun*Transport. Idempotent via the existing class-level marker.

3. Fail-LOUD wire tag (src/nullrun/runtime.py)
runtime.track() now escalates the missing-model warning from logger.warning to logger.error, bumps dropped_llm_call_no_model for dashboards, and tags the wire event with __missing_model: True so the backend's into_track_request gate can reject with HTTP 422 instead of silently recording a zero-cost call. The event is still sent (not fail-CLOSED) so the backend can audit; the flag is wire-private and stripped before persisting.

Files

• src/nullrun/instrumentation/langgraph.py — llm_output-first chain + non-OpenAI sweep
• src/nullrun/instrumentation/auto.py — eager gc sweep for pre-existing Clients
• src/nullrun/runtime.py — ERROR + counter + __missing_model flag
• tests/contract/__init__.py (new package)
• tests/contract/test_llm_call_model_wire.py — 12 tests pinning all three invariants

Test surface
7 unit tests for _extract_model_from_response (langchain-openai 1.x, model key, "any key containing model" sweep, llm_output-before-response_metadata ordering, response_metadata fallback, generations-message fallback, all-empty None + DEBUG log, empty-string fallthrough).
3 tests for track() missing-model handling (ERROR + counter + tag; tag silent when model set; tag silent for non-llm_call types).
2 tests for the eager gc sweep (pre-existing Client gets wrapped; idempotent on re-patch).

Branch base: master @ d4884a7 (after the 0.8.2 version bump from PR #41).

[codecov]

Codecov Report

❌ Patch coverage is 73.17073% with 11 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
src/nullrun/instrumentation/auto.py	62.50%	8 Missing and 1 partial ⚠️
src/nullrun/instrumentation/langgraph.py	85.71%	2 Missing ⚠️

📢 Thoughts on this report? Let us know!