fix(eval): score unknown principals consistently

[vsolano9]

Closes #8.

An unresolvable principal now receives recall 1.0 only when no documents were expected and 0.0 when expected documents are missing. mean_recall also drops the impossible is not None filter, keeping the declared float type and aggregation behavior aligned.

Regression coverage exercises both unknown-principal boundaries against the real PostgreSQL/pgvector fixture.

Validated locally:

• 60 tests on Python 3.12 and Python 3.11 with real pgvector
• both focused regressions repeated 20 times
• strict seeded access-control evaluation: 11 cases, 0 leaks, mean recall 1.0, pass rate 1.0
• wheel and sdist build; clean wheel install, dependency check, import/API smoke, and CLI help
• compileall and git diff --check

Ruff reports no patch-introduced lint or format delta: the single SIM103 finding and existing formatting drift are identical on upstream main.

[github-actions]

Welcome, @vsolano9, and thanks for your first pull request to VaultRAG.

A quick look at CONTRIBUTING.md covers how things are set up and run here.

I'll review this shortly. Ask anything in the thread if something is unclear.