fix(test): repair 6 failing tests blocking develop CI

[ngoclam9415]

Problem

develop CI is red — 6 unit tests fail (the same 6 that fail on PR #18 develop→stable, since #18's head is develop). Develop has been broken since #16 (inject LLMProvider) and the Responses API work (#9/#15) landed.

FAILED tests/unit/core/test_inject_llm_provider.py::test_rlm_resource_set_llm_repoints — ValueError: OpenAI API key not found
FAILED tests/unit/core/test_inject_llm_provider.py::test_ltmemory_set_llm_repoints — ValueError: OpenAI API key not found
FAILED tests/unit/core/test_inject_llm_provider.py::test_set_llm_provider_string_path_updates_all_sinks — ValueError: Anthropic API key not found
FAILED tests/unit/test_llm_providers.py::TestOpenAIReasoningTokens::test_chat_with_reasoning_tokens — TypeError: 'MagicMock' object can't be awaited
FAILED tests/unit/test_llm_providers.py::TestOpenAIReasoningTokens::test_chat_without_reasoning_tokens — TypeError: 'MagicMock' object can't be awaited
FAILED tests/unit/test_llm_providers.py::TestOpenAIReasoningTokens::test_chat_with_zero_reasoning_tokens — TypeError: 'MagicMock' object can't be awaited

Root causes & fixes

1. test_inject_llm_provider.py (3 failures) — tests construct resources with llm_provider="openai"/"anthropic" (string path), which builds a real provider that reads OPENAI_API_KEY/ANTHROPIC_API_KEY from env. They only assert identity wiring and never call the LLM.
→ Added an autouse fixture supplying dummy env keys so construction succeeds offline.

2. TestOpenAIReasoningTokens (3 failures) — thinking models (gpt-5*) now route through the Responses API (client.responses.create) per #9/#15, but these tests still mocked client.chat.completions.create. The awaited responses.create hit an un-mocked MagicMock.
→ Rewrote the 3 tests to mock client.responses.create with a Responses-API-shaped response (output items + usage.output_tokens_details.reasoning_tokens) via a shared _responses_api_response helper. Test intent preserved: verify reasoning-token parsing (150 → 150, absent → None, 0 → None via falsy coercion).

Verification

Full suite, CI-identical invocation (DANA_MOCK_LLM=true DANA_USE_REAL_LLM=false uv run pytest tests/ -m "not live and not deep"):

Before: 6 failed, 1870 passed, 38 skipped
After:  1885 passed, 30 skipped, 166 deselected, 1 xfailed   ✅

Ruff + format clean on both files.

Scope

Test-only change. No production code touched.