Loki Mode

Verification-gate honesty + correctness (6 MEDIUM findings)

Lead theme: gates that silently pass when they should fail (the on-brand-critical class).

• Coverage gate fails loudly on parse drift (.github/workflows/coverage.yml):
  an unparseable coverage line printed a warning and exit 0, silently disabling
  the 70% threshold. Now exit 1 (::error::).
• License audit checks the PINNED version, not @latest (scripts/license-audit.sh):
  the audit queried the bare (latest) package license and used the pinned range
  only as a fallback, so it could pass a non-permissive version that actually
  ships. The pinned/resolved range is now queried first.
• Benchmark patch-validation is no longer a no-op (benchmarks/run-benchmarks.sh):
  has_changes = "+" in patch or "-" in patch was always true (every diff has
  +++/--- headers), so the empty-patch QA gate could never fire. Now counts real
  change lines only.
• Memory retrieval correctness (memory/retrieval.py, token_economics.py):
  detect_task_type crashed on a present-but-None goal/action/phase; skills step
  joins crashed on null/non-str steps; and a token budget threw away task-aware
  ranking (optimize_context re-ranked on raw score, ignoring the computed
  weighted score). All fixed; weighted-score order now preserved under budget.
• emit.sh now writes events.jsonl (events/emit.sh): events emitted via the
  shell helper landed only in the pending dir and were invisible to the
  dashboard (which reads .loki/events.jsonl). Now appends the flat-schema record
  the dashboard consumes.
• get_logs lines param bounded (dashboard/server.py): a negative/zero lines
  value inverted the tail contract; now Query(ge=1, le=10000) -> clean 422.

Each fix ships with a non-vacuous regression test. Full pytest gate: 1176 passed.