Prevent durable PR waits from exhausting execution operation limits #334

Problem

PR-per-unit executions currently poll pull request state with durable timers. A long review wait can therefore consume the durable execution's operation budget even though no application work is happening.

The observed execution reached Lambda's 3,000-operation limit and failed with ExecutionLimitExceeded, while its process row remained RUNNING. During that wait:

• the provider had already merged the reviewed comparison-feature head, but local state still showed the unit and PR as ready;
• the completed views construction stages and draft PR were still valid;
• the watchdog would not notice the terminal durable execution until the configured local expiry;
• the existing repair path would reset completed stages and replay successful construction.

Expected behavior

• A PR review wait persists one callback and does not grow durable operations while provider state is unchanged.
• Provider reconciliation wakes that callback only for merge, closure, reviewed head/target movement, or queued authenticated feedback.
• Concurrent scheduler and feedback wakeups deliver a callback at most once.
• Terminal durable executions are detected while their local process rows are still active.
• Repair reconciles remote PR state, validates merged reviewed-head ancestry, preserves completed stages and artifacts, and resumes an existing draft PR through the current orchestrator alias.
• GitHub, GitLab, multi-repository, and stale-run ownership cases behave consistently.

Proposed implementation

• Add a sparse DynamoDB maintenance index for active executions and parked unit PR waits.
• Replace durable timer polling with persisted callback ownership and an EventBridge-driven provider reconciler.
• Extend the durable watchdog to inspect every indexed active execution.
• Make lane repair provider-aware and state-preserving.
• Add focused regression coverage for long waits, transitions, feedback wakeups, duplicate delivery, early durable failure, and recovery.

Dependency

This work uses the project-level source-control service and authentication model introduced by #333.

Acceptance criteria

• Long unchanged PR waits create no repeated durable operations.
• A merged reviewed head is marked merged only after ancestry validation.
• Completed construction stages retain their state, attempts, and artifacts during repair.
• An existing draft PR resumes without creating a replacement or rerunning construction.
• The repaired execution advances after that PR merges.