[Eval Harness] Step 3-4: Null-agent baseline + run baseline, output scores.jsonl

[Lexus2016]

Parent: #1481 (Eval harness foundation — decomposed per owner request)

This is Step 3-4 of the eval harness (null-agent + baseline run). Depends on #1514 (Step 1: task set) and #1515 (Step 2: runner).

Scope (this issue — ≤200 lines)

Implement the null-agent (an agent that declines every task) and run it against the task set to produce floor baseline scores in scores.jsonl. This establishes the adversarial floor that #1267 requires.

Specific changes

1. Null-agent: a minimal agent config that declines every task (system prompt: "Decline all tasks. Respond with 'I cannot do this.'"). No tools, no reasoning — pure refusal.

2. run_null_agent_baseline(task_set, output_path): runs the null-agent against all tasks in the task set using the runner from [Eval Harness] Step 2: Agent runner — execute task, capture full trajectory #1515, captures trajectories, and writes results to scores.jsonl.

3. scores.jsonl format: one JSON object per line:

   {"task_id": "...", "agent": "null", "tool_calls": 0, "final_answer": "I cannot do this.", "score": 0, "timestamp": "..."}

4. Baseline scoring: a simple scoring function that checks whether the agent made any tool calls toward the expected tools and whether the result matches the expected pattern. Null-agent should score 0 on all tasks.

What this issue does NOT do

• Does NOT define tasks (Step 1 / [Eval Harness] Step 1: Fixed versioned task set (5-10 tasks from evolution cycles) #1514)
• Does NOT build the runner (Step 2 / [Eval Harness] Step 2: Agent runner — execute task, capture full trajectory #1515)
• Does NOT implement hidden-verifier or replay grading ([IMPROVEMENT+REPLACEMENT] Adopt hidden-verifier + continuous partial-credit replay grading (consolidates LHTB + Senior-SWE-Bench) #1302)
• Does NOT run the full agent for comparison (future work after [IMPROVEMENT] Adversarial evaluator floor test — null-agent baseline for evolution metrics (BenchJack) #1267)

Success criteria

• Null-agent config implemented (declines all tasks)
• run_null_agent_baseline() runs null-agent against all tasks
• scores.jsonl output with correct format
• Simple scoring function (0 for null-agent on all tasks)
• Unit test: null-agent produces 0 scores on all tasks
• Unit test: scores.jsonl is valid JSONL
• Fits ≤200-line self-merge cap

Dependencies

• [Eval Harness] Step 1: Fixed versioned task set (5-10 tasks from evolution cycles) #1514 (Step 1: task set) must land first
• [Eval Harness] Step 2: Agent runner — execute task, capture full trajectory #1515 (Step 2: runner) must land first

Unblocks

• [IMPROVEMENT] Adversarial evaluator floor test — null-agent baseline for evolution metrics (BenchJack) #1267 — adversarial floor test (needs null-agent baseline)
• [IMPROVEMENT+REPLACEMENT] Adopt hidden-verifier + continuous partial-credit replay grading (consolidates LHTB + Senior-SWE-Bench) #1302 — replay grading (needs harness + scores format)