tests: cover OpenAI provider gaps — key resolution, Qwen XML salvage, stream edges

[ayobamiseun]

What

Adds tests/test_openai_provider.py (20 tests). The Anthropic and Gemini providers each have a dedicated test module (test_anthropic_provider.py, test_gemini_provider.py), but coworker/providers/openai_provider.py was only covered incidentally — test_providers.py handles complete/stream basics and the param-fix retries, test_provider_router.py handles JSON-form tool-call salvage. This fills what neither reaches:

• Key resolution & client bootstrap (mirrors the existing Anthropic/Gemini tests): resolve_api_key env-over-SecretStore precedence and empty-profile fallbacks; _ensure_client raising the no-key RuntimeError; api_key/base_url plumbed to the SDK client (the Ollama/compat path).
• The Qwen/Hermes XML tool-call template (salvage path 1b, <function=…><parameter=…> — openai_provider.py:296-307), which had no coverage anywhere: happy path per the template in the code comment, verbatim free-text parameters (file content with newlines), unambiguous-JSON coercion, unknown-name filtering, schema-less acceptance, multi-block renumbering, plus _coerce_param directly.
• Salvage gating inside complete()/stream(): salvage fires only when tools were offered and no structured calls came back; structured tool_calls always win over tool-shaped text; streamed text deltas still surface live before the final turn swaps in the recovered calls.
• Stream edge cases: unparseable accumulated arguments surface as {"_raw": …}; parallel calls come out ordered by index with empty args as {}; choice-less keep-alive/usage chunks from compat servers are skipped; and the reasoning attribute variant (xAI/OpenRouter — distinct from reasoning_content) on both stream() and complete().

Tests-only change; no production code touched. Fakes follow the established SDK-free SimpleNamespace pattern from test_providers.py.

Verification

$ .venv/bin/python -m pytest tests/test_openai_provider.py -q
20 passed in 0.54s

$ .venv/bin/python -m pytest tests -q
909 passed, 1 skipped in 42.76s

No screenshots — no UI change.