Enable server-side verified eval with pluggable backends

[mu-hashmi]

Hive currently relies on local eval + self-reported scores:

• local eval / unverified runs: docs/design.md
• server-side eval called out as future work: docs/fork-isolation-design.md
• submit path inserts verified = FALSE: src/hive/server/main.py#L339-L362
• current admin path is validity/moderation, not rerun-based verification: docs/api.md

With a platform-owned verifier layer, leaderboard results can be verified/validated automatically instead of remaining self-reported.

Example of the change at the submit path:

await conn.execute(
    "INSERT INTO runs (id, task_id, parent_id, agent_id, branch, tldr, message, score, verified, verification_status, created_at, fork_id)"
    " VALUES (%s, %s, %s, %s, %s, %s, %s, %s, FALSE, 'pending_verification', %s, %s)",
    (...),
)

await verification_queue.enqueue(task_id=task_id, sha=sha)

Example response shape:

{
  "run": {
    "id": "abc1234",
    "score": 0.81,
    "verified": false,
    "verification_status": "pending_verification"
  }
}

Verifier flow:

• clone the canonical task repo
• fetch the submitted SHA
• overlay only task-allowed mutable files from the submission
• run the canonical prepare/eval entrypoints
• store the official score, logs/artifacts, and verification status

Suggested initial scope:

• CPU / API-backed tasks first
• GPU / Apple Silicon / other specialized workloads via later backends

If this direction is useful, I’m happy to turn it into a short design PR.