Unit tests (T1)

All unit tests pass at 420df50, with OPENAI_API_KEY and TEMPORAL_API_KEY unset: 342 passed, 84 of them new for this work item.

Full unit suite

$ env -u OPENAI_API_KEY -u TEMPORAL_API_KEY uv run pytest tests/unit -q
342 passed in 4.35s

The work item's unit tests

$ env -u OPENAI_API_KEY -u TEMPORAL_API_KEY uv run pytest -v \
    tests/unit/test_config_openai.py tests/unit/models/test_openai_endpoint.py \
    tests/unit/models/test_openai_malformed.py tests/unit/models/test_openai_chat.py \
    tests/unit/service/test_runtime_model.py tests/unit/service/test_o11y.py \
    tests/unit/test_e2e_ollama_guard.py
============================== 84 passed in 3.02s ==============================

File	What it proves	Requirement
tests/unit/test_config_openai.py	base_url / api / context_window_tokens validation, the key rule, abuse cases 2–3, the redactor's input, the shipped Ollama config loads keyless	R1.4, R2, R3.1, R4.3, R6.1
tests/unit/models/test_openai_endpoint.py	the URL a request reaches, OPENAI_BASE_URL ignored, verbatim model, keyless and keyed headers, org/project suppression, redirect refused, model info	R1.1–R1.3, R1.5, R2.2–R2.3, R4.1–R4.2
tests/unit/models/test_openai_malformed.py	unparseable Responses bodies (invoke and stream) → ProviderError	R5.3
tests/unit/models/test_openai_chat.py	Chat Completions mapping, parsing, finish rule, absent fields, streaming assembly, errors	R3.2–R3.4, R5
tests/unit/service/test_runtime_model.py	the startup line's shape; origin drops path and query; a translated error carries no key	NFR observability, abuse case 5
tests/unit/service/test_o11y.py	the chat span's server.address, server.port, tiny_harness.llm.api	NFR observability
tests/unit/test_e2e_ollama_guard.py	the Ollama e2e skips with a reason, never passes silently	T4

Each task's red→green transition is recorded in its commit message (6cead4e, 222cd46, 1879361, bd17981, 420df50).