Loki Mode

Added

• Visible trust trajectory (R4 of the competitive arc): a per-project view of
  whether the agent is earning autonomy over time, derived from real run history
  (council pass-rate, quality-gate pass-rate, iterations-to-completion, and a
  human-interventions axis). The story no competitor tells: you see Loki getting
  more (or less) trustworthy on YOUR repo.
• loki trust (--json): the across-runs trajectory with a per-axis direction
  (improving / regressing / stable, polarity-aware so "fewer iterations" reads
  as improving) and a one-line verdict. Distinct from loki kpis (single-run
  snapshot), which gains a see-also pointer.
• /api/trust/trajectory endpoint + a /trust dashboard panel with sparklines,
  plus a trust_update WebSocket push (mirrors the R3 cost pattern) so the panel
  updates live.

Notes

• Honest data: derived only from .loki/proofs/<run_id>/proof.json; with fewer
  than two runs it says "not enough history yet" rather than inventing a trend,
  it shows a DECLINING trajectory when the data declines (not a feel-good
  always-up chart), and the human-interventions axis is explicitly marked
  unavailable until a per-run intervention counter exists (no fabricated number).
• Council: 3-of-3 unanimous (round 1). bash and Bun share one derivation (the
  bash command and dashboard call the same Python module; the Bun command mirrors
  it byte-for-byte). No duplication of kpis.
• NOT yet wired: the human-interventions axis (lights up automatically once a
  per-run intervention count lands in proof.json).