fix: derive realization narrative from real shared keywords, enforce Tech backbone

The synthesized realization/summary/strategic_goals text was hardcoded
"micro-VM hypervisor" boilerplate unrelated to the actual paired posts.
Now it's built from the real TF-IDF shared-keyword intersection already
computed for each pair, falling back to the two real domain names when
there's no shared vocabulary at all.

Also constrains every cross-domain pair to always include Tech (Hacker
News / arXiv) on one side, per requirement that content keep a
technology backbone across whatever second domain it's paired with.