v0.7.0

A minor rather than a patch: Invocation loses two public properties,
verify() refuses argument combinations it used to resolve by precedence, and
notADouble() gains a required argument. Composer's caret treats that as
breaking on 0.x, and it is — though every one of them was the engine knowing
something and not saying it.

• Invocation::$file and $line are gone. They were promoted readonly
  properties on an @api class that the dispatcher never filled in and
  nothing ever read — public API whose only value was null. Filling them
  would mean a debug_backtrace() on every dispatched call, which this
  package will not pay for a field nobody asked for.

• verify($call, never: true, times: 3) is refused rather than resolved.
  never used to win and the count beside it was discarded without a word, so
  a test could say two opposite things and pass. Both analyser packages report
  this pairing and both promise a runtime counterpart for every complaint they
  make; this is it, in their wording, so a user does not meet a second
  phrasing of a mistake they have already seen.

• verify($call, times: -1) is refused as the nonsense argument it is.
  The bounds now go through Cardinality, which is what the fluent times()
  has always used, so a negative count and a maximum below the minimum are
  caught where they are written. times: -1 used to become the range
  [-1, -1] and fail as an unmet count.

• A facade method handed something that is not a double names itself.
  Understudy::strict(new stdClass()) used to answer "the specification
  closure did not call a method on an understudy" — and strict() has no
  closure, so the reader was sent looking for a mistake they had not made.
  strict, lean, forwarding, label, unused, nothingElse,
  allVerified and transcript each say their own name now, the way
  forget() already did. Knowing which facade was called also separates two
  mistakes that used to share one message: an object that never was a double,
  and a double whose context has been reset — the latter now gets the same
  ForgottenDouble answer a CALL on that object already got.

• A loose double answers : static with itself. The receiver IS the
  double and the generated method really does declare static, so a fluent
  contract chains without a stub for each link. It used to refuse with
  NoDefaultValue — the one return type whose answer needs no invention at
  all. Both READMEs say so, and say why : self is a different claim.

• A return-type conflict no longer prints an unresolved parent. The message
  paths rendered the type without the declaring class, so a reader on PHP 8.3
  or 8.4 was told a target declares : parent and left to work out which
  class that is; 8.5 resolves it in Reflection, which is why it went
  unnoticed. Pinned by a test that fails on 8.3 without the fix.

• Runtime\Mode is @internal, like the rest of Runtime\. It never
  appears in a public signature — strict(), lean() and forwarding() are
  what a user writes — so the @api on it promised a contract nobody could
  reach and no document described.

• Two orphaned docblocks removed, and WhenBuilder says why it is the one
  class here that is not final. FileWrapper::install() says that it only
  ever widens, and that a target list added to the global mode changes nothing
  because global already reaches every class it names.

• A targeted bypassFinals() stops tokenising files it has no interest
  in. The guard was "does the source contain the word final", and
  finally contains it — 58% of a real vendor tree passes, and in targeted
  mode almost none of it declares anything asked for. A file that never names
  a target's class cannot declare that class, so it never reaches the
  tokenizer: 16 µs against 256 µs on a 41 KB file, or roughly a second of
  tokenising per process. The filter is a NECESSARY condition, which is why it
  is safe — its absence is proof, its presence decides nothing and the
  tokenizer still says.

• QueryEquals keeps building its ReflectionMethod per call, and that
  is the answer rather than an omission: caching the decision per class and
  calling the getter directly measures 94 ns against 104 ns, which is 0.6% of
  the 1.65 µs a dispatch through this matcher costs. The number is in the
  code, so the next audit reads it instead of re-asking.

Fixes #95.