Randomness 1.15.0

Added

• |sep: — choose what goes between multiple results. A roll that
  returns several things has always joined them with , , which is
  right inside a sentence and wrong everywhere else. |sep: sets the
  glue: `rdm:3#monster|sep:<br>|link` puts one monster per line,
  `rdm:3#monster|sep:<br>• |link` bullets them, and
  `rdm:3[[Rumours|line]]|sep: /\_` runs them together with
  slashes. It works on every multi-result inline roll — tag rolls,
  table rolls, line and block rolls — and under the dice: prefix.

  Everything after sep: is the separator, spaces and HTML included.
  Escapes cover what the surrounding syntax would eat: \n newline,
  \t tab, \_ space (needed at the end of a glue, since an inline
  span reaches the parser trimmed — sep: / joins with " /", and
  sep: /\_ is the one that gives /), \\ backslash. On a wikilink roll
  sep: goes outside the brackets — inside them a pipe already means
  "column pick" — and on a tag roll it is a segment like unique, so
  it comes before prop:, which still swallows the rest of the line.
  Separator text is never evaluated, so a glue containing [@table]
  prints as itself. Thanks to Gizmo734 for the request.

Fixed

• dice:*|folder=… rolls instead of silently doing nothing useful.
  The tagless * source — "any note matching these properties", with
  no tag constraint — never reached the tag-roll branch under the
  dice: prefix. It fell through to the formula translator and came
  out as the nonsense expression {*|folder=Bestiary|link}: no error,
  no roll, nothing to indicate the syntax was fine and the plumbing
  wasn't. The reference has claimed since these filters landed that
  they work under the compatibility prefix, so this was a documented
  feature that had never run. * now dispatches exactly like #tag,
  including the repetition prefix, |unique, and prop: templates,
  and a test pins the two prefixes to the same note on the same seed.

• Documentation: [@N table] was described as joining results with
  blank lines. It doesn't, and shouldn't — a repeated sub-table call
  runs its results straight together, because it is nearly always
  embedded in a sentence (The party of [@4 hero] sets out.) where an
  injected blank line would be wrong. The separator is the author's
  call, via >> implode: bare for , , or \n / \n\n / <br> /
  ;\_ for anything else. The reference's "Call variations",
  "Filters" and "Repetitions" sections now say so, with a new note on
  the one place separation is automatic — a whole-file MaxReps:
  roll, whose reps are standalone blocks and so are blank-line
  separated. Behaviour is unchanged; the corpus tests now pin every
  one of these joins so the docs can't drift again.

  Two smaller errors went with it: a filter's glue is trimmed and
  never quoted, so the documented implode ", " would have put the
  quote marks in the result and implode , loses its trailing space.
  Both examples are corrected, and \_ is documented as the way to
  end a glue with a space. Thanks to claudermilk for the report.