matcher

Alias-aware slug matching for tool-detection evals.

Wraps verified_aliases.json and an optional secondary-aliases file so a predicted slug counts as correct if it equals the expected slug OR any of the expected slug's verified aliases.

Bidirectional by default: an alias map \{old: new\} is treated as a symmetric equivalence. This matters when the eval set has been renamed to canonical/new slugs but the model under test was trained on the old slugs (or vice versa) - both names should count as the same answer.