Follow-up to #373, from its review. dormant_rules computed one of three distinct diagnoses — reverted / shadowed / excluded — then rendered it away into prose, so the only way to ask which one it was was to match a substring. Four tests did exactly that:

assert "shadowed by 'fix(z)'" in report.undeclared[0][1]

That fails open. Reword a diagnosis for clarity, keep the matched substring by luck, and the test still passes while checking something weaker than it was written to check. A passing test looks identical either way.

The change

_Dormant now carries the distinction the function already computed:

class _Dormant(NamedTuple):
    issue: str
    kind: Literal["reverted", "shadowed", "excluded"]
    detail: str          # the shadower's issue when kind == "shadowed", else ""

main() renders the sentence from a _DORMANT_WHY table. The computation no longer knows any wording, so rephrasing a diagnosis is a change to output alone. kind is a Literal, so mypy rejects a typo'd kind rather than letting it reach a test.

No behavior change, and here is what pins that

The two _run_main tests assert on the rendered EXPLAINED NOTHING lines and are untouched by this PR. They still pass, which is precisely what proves the output format survived a refactor that moved where the prose is built.

Both of #373's mutation proofs still fail as they should — and the first one got stronger:

The swap catching more is the payoff. Under substring matching it was caught only where the two kinds happened to produce visibly different sentences; under kind == every test naming a kind catches it, because the assertion is about identity rather than about text that happens to differ.

Measured

Gates unmoved: 108/0 at 1.4.0, 90/0 at 2.0.0, 1/0 at 2.1.0, all exit 0. Suite 3217, ruff and mypy clean.

🤖 Generated with Claude Code