scaffold v3.32.0 — scaffold run argument passthrough

Makes scaffold run <step> <args…> bind trailing CLI arguments to the step's
$ARGUMENTS placeholder. Previously scaffold run review-pr 376 dropped the
376, emitted an unbound $ARGUMENTS, and agents fell back to raw mmr
commands — losing the Superpowers agent channel, reconciliation, and verdict
logic. Also hardens the tools that consume $ARGUMENTS.

Added

• An "EXECUTE NOW" header prepended to scaffold run output in interactive
  mode (suppressed under --auto / --format json), framing the emitted prompt
  as a runnable workflow and showing the bound arguments — so agents execute the
  step instead of shortcutting to an ad-hoc command.

Fixed

• scaffold run <step> <args…> now binds trailing arguments into the
  $ARGUMENTS placeholder. Previously trailing tokens were dropped and only
  --instructions populated $ARGUMENTS, so scaffold run review-pr 376 emitted
  an unbound $ARGUMENTS and agents fell back to raw mmr commands.
• The $ARGUMENTS substitution now uses a functional replacer, so argument
  values containing $-patterns ($&, $1, ${VAR}) are inserted verbatim, and
  a literal $ARGUMENTS token is never left in the output (it now resolves to an
  empty string when no arguments are supplied — a behavioral cleanup; no pipeline
  step relied on the literal surviving).

Changed

• --fix-threshold=P1 (the = form) is now accepted by review-pr,
  review-code, and post-implementation-review, alongside --fix-threshold P1.
• multi-agent-start / multi-agent-resume prompts now validate the agent name
  to ^[A-Za-z0-9_-]+$ and quote its shell expansions.