The dedup loop only inserted first-seen findings; re-scans skipped existing
fingerprints entirely, so control_refs (re)computed by the mapping passes were
discarded. A finding first seen before control mapping was enabled/tuned would
therefore never gain its control mappings without being deleted + re-added.
Now: existing findings whose re-scan produced non-empty control_refs get updated
in place ($set control_refs). Guarded on non-empty so a run where mapping didn't
run (breakpilot unreachable) can't wipe existing refs. New findings unchanged.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>