v0.3.0

A minor rather than a patch: the misuse rules now fire inside
expectSequence(), where they were silently absent, so a consumer's own code
can draw a diagnostic it did not draw before.

• Allow rasuvaeff/understudy ^0.6. Widened rather than raised: the
  extension works against 0.4, 0.5 and 0.6, and consumers on the older ones
  should not be cut off from it.

• wire()'s shape disagreed with the core on a parameter naming more than
  one contract. An intersection — BookRepository&Auditor — is ONE double
  standing for both, which is what Wire::resolve() builds; the extension
  dropped the parameter, so its key was reported as a missing offset on code
  that wires and runs. A union of two object types makes the core refuse the
  class outright (CannotWire), and reporting a missing key there named the
  wrong mistake: no shape describes a call that never returns, so none is
  produced. Fixes #16.

• The mutation gate rises from 95 to 97. Re-measured on PHP 8.4, the version
  the coverage job pins: 188 of 191 mutants killed, 98.43%, with the same
  three equivalent survivors the config already names. The gate is one mutant
  below the measurement rather than a round number three points under it.

• expectSequence() is recognised as a specification verb. It was known to
  neither spelling's list, so SpecificationRangeCollector never recorded the
  call and understudy.matcherLeak was reported for every matcher inside an
  armed protocol — a false report on code the engine accepts. The closure and
  matcher-kind rules were silently absent on the same closures. Fixes #13.

• VerbNamesTest now walks the core's own public surface and fails when a
  closure-taking verb is missing from either list. expectSequence() was the
  second verb to fall outside every rule after lastCall(); a list nobody
  checks is what let both happen.

• Both READMEs say what was only implied: the understudy.* identifiers are
  stable, because they are what a consumer writes into ignoreErrors, while
  the wording of a message is not.