Loki Mode

Added

• Cost + observability dashboard (R3 of the competitive arc): transparent,
  anti-surprise cost visibility, counter-positioned against the market's #1
  churn driver (surprise AI bills). All built by ENHANCING the existing cost
  surfaces in place, not duplicating them.
• loki cost CLI (--json, --last N): current-run spend, per-model routing,
  per-run history, and budget status, reading the shared
  autonomy/lib/efficiency_cost.py (single source of truth; "not recorded"
  rather than a fabricated $0.00 when no cost data exists).
• New /api/cost/timeline endpoint: intra-run per-iteration cumulative spend +
  per-run history with project totals + read-time budget status, additive to the
  existing /api/cost.
• A self-contained /cost dashboard panel: budget gauge with an 80% warn line,
  model-routing-by-spend, per-run table, inline cumulative chart.
• Pre-cap budget WARNING at 80% (previously only a hard stop at 100%), added to
  the existing check_budget_limit (bash) and checkBudgetLimit (Bun) without
  changing the hard-cap behavior, so an autonomous run flags a budget approach
  before it stops.
• Proactive budget alerting: the budget status now broadcasts over the existing
  dashboard WebSocket loop and shows a persistent banner (amber at 80%, red at
  100%) on every dashboard page, with a poll fallback for late-joining clients,
  so a user not watching the terminal still sees the warning before the cap.

Notes

• Council: 3-of-3 unanimous (round 2, after adding the proactive in-dashboard
  budget banner; round 1 flagged the warning was passive-pull-only).
• No duplication: every cost surface reuses the existing efficiency accounting;
  thresholds (80% / 100%) are consistent across the bash runner, Bun runner,
  dashboard, and CLI.
• NOT tested in this release: the budget banner rendered in every browser engine
  (verified self-contained + reproducible build + endpoint/broadcast tests).