Loki Mode

Defensive hardening from the bug-hunt fleet's lower-severity findings. Each fix
has a regression test proven non-vacuous.

Fixed

• Memory index no longer double-counts on episode re-save. _update_index_with_episode
  de-duplicated episode_ids but incremented episode_count/total_cost_usd/total_tokens
  unconditionally, so a resumed/checkpointed run (same trace id) inflated the
  dashboard Memory panel and loki memory index. The aggregates now increment
  only when the episode id is genuinely new to the topic.
• Memory pattern reads tolerate a corrupt non-dict entry. The pattern loops in
  engine.py (cleanup, get_pattern, find_patterns, index rebuild) now skip a
  non-dict element instead of raising AttributeError and aborting the whole read,
  preserving the per-file _load_json resilience.
• save_pattern no longer loses a save when patterns.json is valid JSON but lacks
  the "patterns" key (partial/external write). It now ensures the list exists
  before the upsert (was a bare subscript that raised KeyError silently).
• Completion council heuristic fallback now requires affirmative positive
  evidence to vote COMPLETE (default flipped from COMPLETE to CONTINUE), so an
  empty .loki/ with no test evidence no longer clears the threshold on
  "absence of failure". Legitimate finished projects (passing or no-test) still
  vote COMPLETE; the change is defense-in-depth behind the existing evidence
  gates. Reuses the same test-results.json signal as council_evidence_gate.
• log_debug writes to stderr, so LOKI_DEBUG=true no longer pollutes the
  detect_rate_limit command-substitution capture (which silently skipped the
  rate-limit reset wait and retried too early into a live limit).