Loki Mode

Added

• Failure-memory loop (default-on, opt out with LOKI_FAILURE_MEMORY=0). Loki
  now learns from its own failures within a run so it stops repeating the same
  mistake. When an iteration fails, the scrubbed crash record (or, when
  telemetry is off, a synthesized record from non-sensitive fields only) is
  mapped to an ErrorEntry on that iteration's episode. The next iteration's
  prompt then carries a "PAST FAILURES TO AVOID:" block built from the most
  recent failure lessons (sorted by true wall-clock timestamp, capped at 3), so
  the agent sees what it just got wrong. Local, zero new setup, zero network.
  Builds on the v7.18.2 crash capture; reuses the existing memory schemas,
  storage, and retrieval. No new scrubbing: every injected value comes from the
  already-whitelisted crash record or non-sensitive fallback fields, so
  docs/PRIVACY.md is preserved.

NOT tested in this release (honest disclosure)

• No live multi-iteration autonomous run validating that the injected lesson
  measurably changes the next iteration's behavior; verification is via the
  module-level loop (crash -> ErrorEntry -> recency read -> injected block) and
  integration tests, not an end-to-end build.
• Lesson usefulness is heuristic. ErrorEntry.resolution is not auto-derived,
  so a lesson states what failed, not a proven fix.
• The cross-run retrieve_anti_patterns secondary rarely matches on local
  setups (documented); the within-run loop is closed by the recency read, not
  that secondary.
• Crash-file-to-episode matching uses the most recent crash file; not tested
  under many rapid same-iteration crashes.
• The Bun route build_prompt is an intentional stub; this is a bash-route
  feature with prompt parity preserved (no Bun change).