The `harness2-judge`'s grade persistence logic during task re-runs and judge failures, specifically the `_merge` operation's handling of run IDs and rollouts.

## Description

If a task is re-run and the next `harness2-judge` call fails for that task, `judged.json` keeps the grade of the rollout that was replaced. The report then shows a score for a deliverable that no longer exists.

## Problem

Run ids are fixed labels (`//`), so a re-run task writes its new rollout under the same run id. A partial task re-runs on resume, for example. `_merge` takes "same run id" to mean "same rollout". When the new judge call fails, it keeps the old graded row in place of the new ungraded one.

This repro uses `DemoAdapter`, a judge that grades the answer on disk, and the judge CLI's own collect / judge / merge / write steps. No network is used:

```
attempt 1: partial # sequential mode hit a transient 429
judge 1: demo-oc-agg1/ledger-1/agg1 1/1
attempt 2: ok # resume re-runs the task; the new rollout is wrong
on disk: demo-oc-agg1/ledger-1/agg1 -> 'Total: 105.25'
judge 2: demo-oc-agg1/ledger-1/agg1 1/1 graded=True # judge was down; old 1/1 kept
```

Repro script

```python
import json, tempfile
from pathlib import Path
from unittest.mock import patch

from harness2.cli import judge as JCLI
from harness2.core import judge as J, recursion, schedule
from harness2.core.proposer import Reply
from harness2.demo import DemoAdapter, DemoProposer

class Adapter(DemoAdapter):
judge_down = False

def score(self, run_...