docs: [ENG-2884] expand SKILL.md §7 with sample envelopes, thresholds, worked example

A skill-discovery sub-agent test surfaced gaps in §7 (brv dream):
prose described the surface but no sample JSON envelopes existed,
prune-signal thresholds were left to the agent to guess, and there
was no end-to-end walkthrough — so consumers had to grep the source
to learn field shapes.

Additions to §7:
- Sample dream scan --format json envelope (sessionId + candidates
  with one example each of link / merge / prune / synthesize shapes,
  including signals + daysSinceModified for prune).
- Sample dream finalize response (archived, skipped, logId, status).
- Prune signal thresholds inline: importance < 35 → low-importance,
  draft >60d / validated >120d → stale-mtime, maturity 'core' never
  surfaced.
- 'Worked example — prune the stalest topic (no curate detour)':
  three-block scan → finalize → undo walkthrough.
- Path-attribute warning in the merge bullet so agents stripping the
  scan-emitted .html before <bv-topic path=> see the guidance.
- Updated undo description to reflect the symmetric mtime + signal
  restoration shipped in this PR — was 'mtime resets to current time';
  is 'content, original mtime, and sidecar runtime signals... all
  restored'.

Companion change in prune-candidates.ts: sync-hint comment near the
threshold constants pointing at SKILL.md §7, so future threshold
changes don't silently drift the agent-facing docs.