fix: critical bugs in compress_research, LangGraph deprecations + first unit test suite (77 tests)

[at384 (zamal-db)]

Summary

This PR fixes 3 confirmed bugs and adds the first-ever unit test suite (77 tests) for open_deep_research.

---

Bug Fixes

1. Wrong model in compress_research token-limit check (relates to #230)

File: src/open_deep_research/deep_researcher.py L569

The compress_research function invokes configurable.compression_model (L528) but the is_token_limit_exceeded fallback was checking against configurable.research_model. When a token-limit error is raised by the compression model, the check against the research model can silently fail to match the provider causing the loop to raise instead of gracefully trimming messages.

`diff

• if is_token_limit_exceeded(e, configurable.research_model):

• if is_token_limit_exceeded(e, configurable.compression_model):
  `

2. remove_up_to_last_ai_message removes newest context instead of oldest (relates to #252)

File: src/open_deep_research/utils.py

The original implementation searches backwards (range(len-1, -1, -1)) and returns messages[:i] removing the most recent AI exchange (the freshest research). The function is called in the compress loop to shed context when hitting token limits, so it should remove the oldest exchange first.

Fixed: Forward search finds the first AIMessage skips its associated ToolMessage replies returns everything after progressively dropping the oldest research while preserving recent context.

3. LangGraph V05 deprecation warnings (relates to #225)

File: src/open_deep_research/deep_researcher.py L591, L703

StateGraph(output=...) and StateGraph(input=...) are deprecated since LangGraph V05:

`diff

• output=ResearcherOutputState,

• output_schema=ResearcherOutputState,

• input=AgentInputState,

• input_schema=AgentInputState,
  `

Note: config_schema= is intentionally left unchanged it maps to context_schema= in V10 which changes runtime behavior and should be migrated separately.

---

Improvements

Extended provider support in get_api_key_for_model

Replaced the hardcoded 3-provider if/elif chain with a data-driven _PROVIDER_API_KEY_MAP covering 9 providers: OpenAI, Anthropic, Google, Groq, DeepSeek, Mistral, Cohere, Fireworks, and Bedrock (AWS). Adding a new provider is now a single dict entry.

---

First Unit Test Suite (77 tests)

This repo had zero tests. This PR adds a comprehensive tests/unit/ suite:

Module	Tests	Covers
test_utils.py	56	remove_up_to_last_ai_message, is_token_limit_exceeded, get_model_token_limit, get_config_value, get_today_str, anthropic_websearch_called, openai_websearch_called, get_api_key_for_model, get_notes_from_tool_calls
test_configuration.py	11	SearchAPI, MCPConfig, Configuration (defaults, from_runnable_config, env overrides)
test_state.py	10	override_reducer, all structured output models (ConductResearch, ResearchComplete, Summary, ClarifyWithUser, ResearchQuestion)

All 77 tests pass. No new dependencies added.

---

Verification

• All 77 unit tests pass (uv run pytest tests/unit/ -v)
• Integration tested end-to-end with live Vertex AI Gemini API (graph execution: clarify_with_user write_research_brief research_supervisor)
• git diff --stat: 7 files changed, 827 insertions(+), 22 deletions(-)