v2.0.0 — Clean refactor

A from-scratch rewrite. The v1 CLI (needlehaystack.run_test) is gone;
the new entry point is niah. Existing v1 result files preserved in
original_results/.

Highlights

• New niah CLI: run, validate, reconstruct
• Two-YAML config model: one model spec + one run spec
  (replaces v1's 25-flag invocation)
• New tasks: single, multi, uuid, uuid_chain (multi-hop UUID
  chain for testing multi-step long-context reasoning)
• Modular extension points via small Protocols + registries — add
  a task, provider, haystack, or scorer without touching the runner
• JSONL result store with recipe-based reconstruction: rebuild the
  exact context a model saw without storing megabytes per row
• uv-based packaging, Python 3.12+, 207-test suite with FakeProvider
  (zero API keys needed to develop)
• Bug fix: multi-needle depth-percent reporting

Install

pip install needlehaystack==2.0.0
# or: uv add needlehaystack==2.0.0

Upgrade from v1

# before:
needlehaystack.run_test --provider openai --model_name gpt-4o ...

# after:
niah run configs/runs/single_needle.example.yaml

See the README quick-start for the YAML config shape.