test(scoring): pin the empty-forbidden and empty-appropriate rate conventions

[dchaudhari7177]

Closes #38.

score() defines disclosure-rate as 0.0 when forbidden is empty and utility as 1.0 when appropriate_flows is empty. Neither had a direct test, so a refactor could divide by zero or flip a default with CI green.

New tests/test_rate_conventions.py — stdlib only, deterministic, no network:

test	asserts
empty forbidden	disclosure-rate exactly 0.0, no ZeroDivisionError
empty appropriate_flows	utility exactly 1.0
both empty	(0.0, 1.0) — the degenerate scenario
naive agent, built-in suite	aggregate disclosure-rate non-zero (1.0, and 1.0 per scenario)
compliant agent	the other end of the same axis, (0.0, 1.0)

The first three are parametrized over four output shapes — silent, discloses to the appropriate recipient, broadcasts to everyone, says nothing useful — because independence from the outputs is the entire content of the convention.

The naive-agent test is the discriminating case the issue asks for. Without it, a scorer that returned the empty-matrix defaults on every input would satisfy every other test here while measuring nothing at all.

Mutation-checked

A convention test that passes against a broken implementation is worthless, so:

• flipping both defaults (0.0 -> 1.0 and 1.0 -> 0.0) fails the degenerate-scenario cases;
• removing the empty-forbidden guard raises ZeroDivisionError inside the first test rather than passing quietly.

Test-only — scoring.py is untouched. 80 passed; ruff check, ruff format --check, mypy src clean.