Accept an IndependenceBasis member as itself in _basis_of

[Silentpartnercoding]

One branch plus four regression tests.

The bug

_basis_of resolved the basis with IndependenceBasis(str(raw)).

IndependenceBasis mixes in str — but Enum.__str__ wins, so str() on a member returns "IndependenceBasis.ATTESTED", not "attested". Lookup raised ValueError and the basis fell through to UNKNOWN.

So a caller passing the vocabulary's own members — the natural, type-correct call — had every basis silently downgraded.

What it did and didn't affect

It failed in the conservative direction: no verdict was ever more confident than the evidence allowed. And nothing recorded is affected — every caller in this repo annotates the field str | None and passes strings.

It still mattered. attested_margin and the ATTESTED path in asymmetric_verdict key off this value. A downgrade doesn't merely lose a label — it removes a root from the attested count, which is an input to a decision, not a description of one. An attested counterexample handed in as an enum member could read INDETERMINATE where REFUTED was correct.

The fix

A member is returned as itself. The string path is unchanged and remains the normal one. Unrecognised values — wrong case, near misses, empty string, non-strings — still read as UNKNOWN.

Tests

Four new, all of which fail without the branch (verified by removing it and re-running: 4 failures). One pins str(IndependenceBasis.ATTESTED) != "attested" so the branch isn't later deleted as redundant by someone who reasonably assumes a str-mixin stringifies to its value.

full suite   247 tests, OK

Found while building the NIST independence fixture (#107), where passing the enum produced weakest basis: unknown for claims explicitly marked declared.

🤖 Generated with Claude Code