feat: SWE-bench submission format export

[christso]

Objective

Add an export adapter that converts agentv results to the SWE-bench submission format, enabling participation in the SWE-bench leaderboard.

Motivation

AgentV's Docker workspace feature (#971) makes it capable of running SWE-bench evaluations. However, to submit results to the official SWE-bench leaderboard, results must be in their specific format (all_preds.jsonl + metadata.yaml + trajs/). A built-in exporter bridges this gap.

Design

Following "Lightweight Core, Plugin Extensibility" — this should be a CLI command or export adapter, not a core feature.

Proposed interface

agentv export --format swe-bench <results-dir> --output <submission-dir>

Output structure

submission/
├── all_preds.jsonl          # {instance_id, model_name_or_path, model_patch}
├── metadata.yaml            # Agent scaffold description
├── README.md                # Auto-generated from config
├── trajs/                   # Agent trajectories
│   └── <instance_id>.json   # From agentv trace data
└── results.json             # Converted from index.jsonl

Mapping (agentv → SWE-bench)

AgentV Field	SWE-bench Field
test_id	instance_id
unified_diff / output	model_patch
target	model_name_or_path
trace_summary	trajs/<instance_id>.json
score, scores[]	results.json

Acceptance Criteria

• agentv export --format swe-bench produces valid SWE-bench submission directory
• Generates all_preds.jsonl with correct instance_id → model_patch mapping
• Generates trajectory files from trace data
• Auto-generates metadata.yaml from agentv config
• Can be submitted to SWE-bench leaderboard validation

Non-goals

• Auto-submitting to leaderboard (just generate the format)

[christso]

Closing as YAGNI — the SWE-bench export format is only needed to submit to the leaderboard, which isn't a near-term goal. The import command (#978) covers the actual use case of running SWE-bench evals through agentv. If leaderboard submission becomes a priority later, this can be reopened.