Loki Mode

Memory data-integrity fixes + dashboard UX

Two HIGH memory data-integrity fixes and three dashboard usability improvements.
Council 3/3. Memory: 28 + 9 existing tests pass plus adversarial concurrency
probes (200-800 threads, zero lost updates).

Memory:

• Lost-update fix in pattern usage counting: increment_pattern_usage now performs
  the entire read-mutate-write inside one exclusive file lock (new
  storage.increment_pattern_usage, mirroring the boost path). The old path read
  under a shared lock released immediately, mutated a detached object, then wrote
  it back wholesale, dropping concurrent increments. The lock is reentrant, so the
  inner atomic write does not deadlock.
• Ingest idempotency: episode ids are now derived deterministically from the stable
  session/task key (ep-<sha1[:12]>) with an existence check, so re-ingesting the
  same session no longer double-counts episodes / cost / tokens. Keyless ingests
  keep a random id (no false aliasing).
• Cross-project pattern JSONL append is now done under an exclusive flock with
  flush + fsync (buffer-all-rows-first), so concurrent writers cannot interleave or
  tear lines; the reader logs dropped corrupt lines instead of silently skipping.

Dashboard:

• Overview tasks list: pagination + compact-by-default cards + status/priority
  filtering, replacing the unbounded infinite scroll.
• App Runner: the embedded app preview uses the full available width, and when the
  running app exposes more than one service (UI plus API) the preview shows
  per-service tabs. The tab bar is latent for single-service apps. The iframe src
  is restricted to http/https and all labels/URLs are escaped.
• Sidebar: running apps are surfaced in a dedicated, emphasized group (dropdown +
  live count + tidy stop list), visually separated from inactive/known projects,
  replacing the clumsy flat pill stack. Focus/stop/poll behavior is unchanged.