orcho-mcp 0.8.2

Two defects that made a paused or abandoned run unreadable from the client
side, both found in a field report against 0.8.5.

Changed

• Requires orcho-core 0.9.0. That release makes the dead-run detection this
  server relays actually fire: the liveness predicate behind stalled could
  never become true on a real run, so orcho_run_diagnose answered active
  for runs whose processes had been gone for hours.

Fixed

• A paused verification gate now tells the operator what was rejected. The
  engine writes its findings — severity, evidence, and the required fix —
  under meta.phase_handoff.artifacts.findings; this server read only the
  payload's top level, so every gate pause arrived as verdict: REJECTED with
  findings_summary: null and nothing to act on. The reader now consults both,
  and a contract test drives it from the engine's own payload builder rather
  than a hand-written fixture, which is how the mismatch survived.
• A run abandoned by a dead server process is settled instead of being
  reported as live work forever. Runs are detached children reaped by a task
  inside this process; when the process itself goes away — a client restart, or
  a kill that takes the tree with it — nothing is left to record that the run
  ended. The recovery probe that exists for exactly this was never called
  outside the test suite. It now runs once per server start, before the first
  client request, and marks orphaned only those running entries whose
  recorded pid is proven dead. A probe failure is reported on stderr and never
  blocks startup.

---

Install: pip install orcho-mcp==0.8.2 — https://pypi.org/project/orcho-mcp/0.8.2/