Review task: don't update last_seen_review_id when LLM response fails

[ClaydeCode]

Problem

When addressing a PR review, if the LLM invocation fails (e.g. due to an API overload/overloaded_error), the CLI backend returns an error JSON instead of raising `UsageLimitError`. The `parse_response` call then raises a `ValueError`, which is caught and logged, but the code continues and still updates `last_seen_review_id` to mark the review as seen. The review is effectively dropped.

This was observed for FreeshardBase/freeshard-controller#160 (PR #163):

Failed to parse address_review response JSON: LLM response failed validation for AddressReviewResponse: 1 validation error for AddressReviewResponse
  Field required [type=missing, input_value={'type': 'error', 'error'...}, input_type=dict]
Data: {'type': 'error', 'error': {'type': 'overloaded_error', 'message': 'Overloaded'}, 'request_id': 'req_011CZqrNqKPPVeYVfV5ZseM3'}

Despite the failure, the log then shows "Review comments addressed" — which is incorrect.

Root Cause

In tasks/review.py, the except ValueError block (lines 133–136) falls through to the state update that sets last_seen_review_id. This should not happen when the failure is due to an LLM/API error rather than a minor formatting issue.

Fix

In the except ValueError block, check if the error output looks like an API error (e.g. {'type': 'error', ...}). If so:

• Log at error level
• Return early without updating last_seen_review_id
• Optionally set status to a retryable state (e.g. stay at pr_open)

This ensures that on the next turn, the review is picked up and addressed again.