Loki Mode

App Runner shows running apps + URLs, intelligent tasks, default-on advisory gates

• App Runner now detects externally-launched apps (compose stacks) and shows
  the service URL (dashboard/server.py, dashboard/static/index.html via
  dashboard-ui/): previously the panel showed "Not Started / No app yet" even
  when a full docker-compose stack was running (e.g. web on :3000 + api on :8000),
  because the app-runner only recorded state for apps IT launched. The dashboard
  now discovers a running compose stack for the active project (bounded, fail-open
  docker probes; ~2.5s cache; offloaded off the event loop), identifies the
  primary web service, and surfaces a clickable URL + a live-preview iframe with
  honest status (running/starting). Never fabricates a URL for a non-running
  container. autonomy/app-runner.sh also persists url/primary_service/runtime
  port for its own compose launches so both paths look consistent.
• App Runner UI polish: one panel instead of three redundant headings, fixed
  the right-edge content clipping, friendlier empty states, clearer "Agents
  running / Tasks queued" labels, sidebar name tooltips.
• Intelligent task enrichment (autonomy/run.sh, autonomy/lib/prd-enrich.sh):
  PRD-derived tasks no longer show placeholder junk (description == title,
  "...so that the product delivers its core value" boilerplate). A deterministic
  fix routes the real spec section body into the description, and an LLM
  post-pass synthesizes a real description, acceptance criteria, and user story
  grounded in the spec. Degrades gracefully to the deterministic output when the
  provider is unavailable; never blocks queue population, never fabricates.
• Default-on advisory verification gates: LSP diagnostics, semantic
  test-authenticity, and invariant/property gates now run by DEFAULT (advisory
  surfacing into the next iteration's prompt), extending the mock/mutation
  precedent, on both routes. Blocking is opt-in via *_BLOCK flags. A clean run
  never newly blocks (deny-filtered); coverage measurement stays opt-in.
• Honesty fixes: removed the CodeQL claim from the static-analysis gate docs
  (we run ESLint/Pylint/type-checkers, not CodeQL); corrected the Codex CLI flag
  in README (--sandbox workspace-write, the deprecated --full-auto replacement).

Gates: local-ci 83/83, bash/Bun parity 6/6, 97 Bun gate tests + full pytest
green, 3-reviewer council unanimous APPROVE (after rounds that correctly caught a
built-vs-source divergence and stale docs, both fixed).