v2.82.1 — Fix: legal & insurance archetype detection (found by pipeline testing)

Fix: legal & insurance archetype detection (found by pipeline testing)

• legal and insurance projects were misclassified as library. Their
  archetype score() functions check raw README terms (iolta, matter,
  acord, naic, underwriting…) that mineReadmeKeywords() never emitted, so
  detection always scored ~0 from README. Insurance was latently broken; legal
  shipped this way in 2.81.0. Fixed by emitting the legal + insurance detection
  terms from mineReadmeKeywords() (bare claim deliberately excluded as too
  generic). A real legal project now detects as legal (high confidence), a
  real insurance project as insurance — so init writes the correct reviewer
  set + gates and ci archetype-drift is accurate.
• Regression coverage these verticals shipped without: added
  tests/fixtures/legal-clio and insurance-acord; run-archetype-e2e is now
  31/31 (was 29). Runtime auto-attach was already correct for all reviewers —
  this fixes the init-time classification path. Full suites 336 + 268 green.

---