fix(linear): orchestration reconciler loses terminal events during outages (no stranded-orchestration backstop)

[isadeks]

Component

API or orchestration

Describe the bug / gap

The #247 Mode A orchestration reconciler reacts to TaskTable-stream terminal-state events to release dependency-unblocked children. If the reconciler is unavailable when a child reaches terminal state (deploy window, throttle, a crash like the 256MB OOM found during the first dev smoke), that stream event is lost and never reprocessed — the dependent children are never released and the orchestration stalls forever with no recovery.

Observed live on dev (2026-06-09): child A reached COMPLETED during a reconciler OOM window; after the fix deployed, A's completion event was gone, so B stayed blocked until the record was manually nudged to re-emit a stream event.

Proposed solution

Add a scheduled stranded-orchestration reconciler (mirrors the existing reconcile-stranded-tasks pattern): a periodic sweep that scans OrchestrationTable for orchestrations with children in non-terminal states whose predecessors are already terminal-success, and releases them (idempotent — releaseChild is already idempotency-keyed). Also surfaces orchestrations stuck with a released child whose task is long-terminal but whose row never advanced.

Acceptance criteria

• A child completing during a reconciler outage is eventually released (bounded by the sweep interval).
• Sweep is idempotent — no duplicate child tasks, no regressed statuses.
• Stuck/stranded orchestrations are logged/metriced for operator visibility.

Related

• feat(linear): parent/sub-issue orchestration with dependency-aware child tasks #247 (Mode A orchestration — introduces the reconciler this backstops)
• Mirrors reconcile-stranded-tasks.ts + StrandedTaskReconciler construct.