Loki Mode

Memory integrity + checklist/cost trust fixes + dashboard reuse (wave-7 hunt)

• Anti-pattern memory (memory/retrieval.py): anti-patterns were
  double-counted when present in both patterns.json and the legacy
  anti-patterns.json (now deduped); search_all_namespaces omitted the
  anti_patterns collection entirely, so "do not repeat this mistake" memories
  were silently missed in cross-namespace search (now included); vector-index
  staleness was checked only for the semantic collection after consolidation
  (now also the anti_patterns collection invalidates over its source files).

• Dashboard no longer killed by a build (autonomy/run.sh): when a build
  reclaimed the dashboard port (the defense-in-depth port-reclaim path), it killed
  the listener on that port, which could drop a user's own live dashboard mid-use
  (connection refused, websocket failure). It now probes /api/status and REUSES
  a healthy dashboard already serving on the
  port instead of killing it (opt out: LOKI_DASHBOARD_FORCE_RECLAIM=1).

• Checklist no-fake-green (autonomy/checklist-verify.py): the tests_pass
  check ran with --passWithNoTests, so a zero-match test pattern reported
  SUCCESS - a checklist item that REQUIRES test verification passing with no
  tests run. It now requires at least one test to be discovered.

• Cost aggregation glob (dashboard/server.py): cost totals globbed
  *.json (including non-iteration files); now uses iteration-*.json,
  consistent with the budget-snapshot and cost-timeline readers.

  Found by the wave-7 adversarial hunt.