Runtime: add per-task retry budgets and escalation policy

[jafreck]

Problem

• Some tasks spend long wall-clock time in repeated code-migrator/parity/failure-recovery cycles before ending blocked.

Proposal

• Add configurable per-task budgets (max attempts, max wall-clock, max token spend).
• Introduce deterministic escalation after budget thresholds (e.g., reduced scope prompt, alternate model tier, or explicit manual intervention marker).
• Surface budget burn-down in progress logs.

Acceptance Criteria

• Tasks stop infinite churn after budget thresholds.
• Escalation path is deterministic and observable.
• Progress/checkpoint includes budget usage fields.
• Add tests for budget exhaustion and escalation behavior.

Context reference

• Long-running loops with many 300-390s heartbeat windows observed before eventual block.