adversary: unbounded tail-verification loops cause dispatch timeouts — codify execution-budget discipline

## Observed
During a per-story adversarial convergence arc (8 passes), two consecutive adversary dispatches died on API request timeouts, both mid "let me verify one more thing" tail-loops after their substantive review was complete. A third dispatch with an explicit execution-discipline preamble ("one spec read-through, one code sweep, one enumeration pass, then WRITE THE REPORT; open questions become report caveats, not verification rounds") completed reliably — as did all five subsequent passes carrying the same preamble.

## Engine gap
The adversary agent prompt has no pacing/budget discipline. Fresh-context full-enumeration reviews (every PC/POST/EC/AC) naturally accumulate long single turns; without an explicit budget the model spirals into unbounded tail verification, exceeding request timeouts and losing the entire review (the report was never emitted).

## Suggested fix
Add to the adversary agent prompt: an execution-budget protocol (bounded sweep phases, then mandatory report emission; residual uncertainty is recorded as explicit report caveats with severity rather than triggering additional verification rounds). Optionally: emit the report incrementally or checkpoint findings so a timeout doesn't lose completed work.

## Evidence
Session 2026-07-06/07, private consumer project: passes 2a/2b died (600s+ stalls, request timeouts), pass 2c + passes 3-8 with the bounded preamble all completed in 7-22 min.