A build can wedge permanently on repeated identical escalations; rescue can't recover

[n1arash]

Summary

A build can wedge permanently: an issue re-escalates with the same reason every rescue cycle, and the auto-reviewer keeps re-queuing the identical rescue with no progress — burning cycles and tokens, and blocking dependent issues.

Evidence (dogfood campaign)

• F1/ISS-001 escalated (turn budget, then "repeatedly finished without a progress.md handoff").
• The auto-reviewer rescued it 4 times through the attention queue; it re-escalated each time and ended needs_human.
• This blocked 3 dependent issues → feature build:stuck.

Why it matters

Unattended, this burns rescue cycles and tokens with no forward progress, and silently blocks the rest of the feature.

Proposed fix

• Detect a repeated identical escalation reason and stop re-queuing the same rescue.
• Escalate to a different action instead — e.g. "split this issue" or "raise its budget" — rather than looping the same answer.
• Root cause ties to the turn-budget issue (the worker never reaches the handoff before the turn cap), so this is a safety net on top of that fix.

Acceptance criteria

• Repeated identical escalation reasons are detected (N consecutive same-reason escalations).
• After the threshold, the system stops re-queuing the same rescue and surfaces a distinct "split / raise budget" action.
• Dependent issues are not silently blocked without surfacing the wedge to the operator.
• Regression test simulates a same-reason escalation loop and asserts it breaks out.

Source: dogfood/ITERATION_REPORT.md MAJOR-5; dogfood/TIMELINE.md build_round_0–3.