Loki Mode v7.104.3

Task list accuracy: honest, stable done column (no empty cards, no duplicates)

• The dashboard "done" column now shows honest, populated iteration cards.
  Previously, when an iteration finished, track_iteration_complete() (autonomy/run.sh)
  wrote a THIN completed marker ({id, type, title: "Iteration N", status, exitCode, provider}) with no description and no logs, and then deleted the RICH in-progress
  record - so completed iterations rendered as empty cards with only "Iteration N"
  and nothing inside. The completer now LIFTS the honest body (logs, startedAt) from
  the in-progress record before removing it, and writes an honest, iteration-scoped
  title ("Iteration N complete - " / "Iteration N failed (exit )") plus a
one-line description with the phase and duration. It deliberately does NOT reuse the
borrowed PRD-story title from the in-progress record (that title is always the first
pending story, so carrying it onto N done cards would have falsely implied that story
was built N times - a fake-green the trust model forbids).

• No more duplicate iteration cards. The completer replaced a blind
  data.append(...) (which accumulated the same id across sub-runs - real data showed
  iteration-1 five times) with an upsert-by-id, and now makes the completed/failed
  terminal state mutually exclusive per id (an id is removed from the other terminal
  file and from in-progress on completion).

• GET /api/tasks deduplicates by id with a terminal-wins rule (dashboard/server.py).
  The dashboard reads dashboard-state.json task groups first; against pre-fix or
  partially-written state the same id could appear in more than one column (observed:
  iteration-13 in in-progress AND completed AND failed at once). The endpoint now
  keeps exactly one entry per id, preferring the terminal (done) record so a completed
  iteration no longer lingers in the in-progress/todo columns. Pending PRD-story ids and
  iteration ids never collide, so no task that is legitimately pending is dropped.

• Regression coverage: tests/test-iteration-complete-accuracy.sh (extracts and drives
  the real run.sh completer block - honest title, lifted logs, upsert, mutual exclusion)
  and tests/dashboard/test_tasks_dedup.py (terminal-wins dedup across columns).