review-batch: single-review JSON failure aborts entire session

[don-petry]

Problem

During the scheduled review run (run #25707852006), the single-review step invoked claude-opus-4-7 to confirm an approval for PR #129. The model ran for ~74 seconds and returned a non-JSON response, causing the entire batch session to abort — skipping all 35 remaining PR candidates.

Failure trace

[approve] triage cleared — running single confirmation (claude-opus-4-7)
##[error]single-review did not produce valid JSON
##[error]Review failed for .../pull/129 (exit code 1)
##[error]Session aborted early after failure on .../pull/129 (exit code 1). Skipped 35 remaining candidate(s)
Summary: 0 reviews posted, 0 no-ops skipped, 1 failures (processed 1/36 candidates) [SESSION ABORTED EARLY]

Root cause

The audit/confirmation model (claude-opus-4-7) did not return structured JSON — likely due to a verbose non-JSON response, a refusal, or a timeout/context-limit issue. The scripts/review-batch.sh script treats any per-PR failure as fatal and exits, aborting all remaining candidates.

Acceptance Criteria

• Resilient JSON parsing: If single-review (or any review engine call) returns non-JSON output, the error is caught, logged with the raw output, and treated as a per-PR failure — not a fatal crash.
• Per-PR isolation: A failure on one PR does not abort the session. Remaining candidates continue to be processed.
• Retry logic: The single-review step retries at least once (with a short delay) before marking the PR as failed, to handle transient model errors.
• Fallback behavior: If the audit model fails after retries, the system can optionally fall back to the triage-level decision (or skip and flag for manual review) rather than hard-failing.
• Accurate summary: The final session summary correctly reports the number of skipped/failed PRs and does not claim SESSION ABORTED EARLY when only individual PRs fail.
• No regression: Existing review logic (triage → escalation → deep review → audit flow) continues to work correctly for PRs that do not encounter errors.