Summary

Wires RCV prompt generation into the prerender audit pipeline. The default audit behaviour
is completely unchanged — prompts are generated only when generatePrompts=true is explicitly
passed (opt-in). Also plumbs LLMO site config (categories/topics/regions) through to Mystique
so generated prompts can be correctly classified.

What changed

src/prerender/handler.js — sending the Mystique request

• sendPrerenderGuidanceRequestToMystique accepts a new generatePrompts parameter
  (defaults to false) and includes it in the SQS payload
• Each suggestion now carries hasPrompts: true when it already has stored prompts, so
  Mystique can skip re-generation
• Reads LLMO site config (categories, topics, region codes) from
  site.getConfig().getLlmoConfig() and sends them as llmoCategories, llmoTopics,
  llmoRegions
• handleAiOnlyMode parses generatePrompts from the SQS data field using the same
  safe JSON parsing pattern already used for opportunityId and scrapeJobId

src/prerender/guidance-handler.js — receiving Mystique's response

• Destructures prompts from each incoming suggestion alongside url, aiSummary, valuable
• Persists prompts with a preserve-or-update strategy:
  prompts: hasNewPrompts ? prompts : (currentData.prompts ?? [])
  New prompts win; existing prompts are kept when Mystique returns an empty array
• Adds suggestionsWithPrompts and totalPromptCount to the quality-metrics log line