v0.5.16: chore: 0.5.16 (#200)

• Drag and drop. A run can drag one control onto another, such as a card onto a board column, when the page
  offers both ends. A drag is checked like a click before it is made: one that would commit something that
  cannot be undone stops for confirmation unless the run is authorized, and a drop that changes nothing sends
  the run to recovery instead of being repeated. Both ends are rechecked the moment before
  pressing, and a drag whose target changed is decided again from the new page instead of being retried. A drop
  target that changes while the card is in the air is not dropped on: the card is carried back and let go where
  it was picked up. A drag between a page and a frame hosted separately from it is refused, as is a drag onto
  its own source.

• The fixture eval suites (local and mock) run on a schedule and on demand in .github/workflows/evals.yml, and
  the job fails on any regression, so a change to browsing behaviour that breaks a task no longer waits for
  someone to run the suite by hand. The job reports a per-task pass rate over its repeats and refuses a run
  where any task is below the target.

• An authorized run can submit a form after attaching a file. Unsure whether the submit was the commit the
  task meant, it asked recovery, and was then refused the same click again whatever recovery said, until it
  ended stuck with nothing sent. When recovery, shown the page and the task, names the same control that was
  refused, the run may now make that click once. A credential change is still never let through this way, and
  a run without authorization still stops for confirmation.

• The stateful mock suite grows from 20 tasks to 26: a story opened in a second tab, a native date input, a
  newsletter form with a hidden field a careful run leaves alone, an export downloaded from behind a sign-in,
  a document attached to a form and submitted, and a card dragged from one board column to another. A run may now carry attachments, the way a caller supplies
  a file.

• Answer grading folds case and typographic punctuation before it matches, so a run that words a right answer
  differently is not failed for its phrasing, and the runner warns when a task uses more than 60% of its step
  budget.

• docs/evals.md states how to read a published number, including the caveats that have to travel with one.

• A run asked for something the page does not have, such as the heading of a page with none, now stops after
  one recovery instead of spending all of them. Recovery used to answer "report that it is absent and finish",
  the completion check refuses any finish while requested information is unread, and the two repeated until the
  run ended stuck. A recovery that can only direct such a finish now ends the run with its diagnosis as the
  error.

• A stuck run returns what it did read. Its answer carries the cited facts gathered before it stopped, marked
  as partial, with their evidence and citations, so a caller no longer has to browse again for them.