Dev: SPOT5 benchmark + codebase restructure → main

[Mtrya]

Status

Draft PR for ongoing development on the dev branch.

Overview

This branch implements foundational infrastructure for AstroReason-Bench - a
benchmark suite for evaluating LLM agents on space mission planning problems.

Current Changes (ready for review/testing)

• ✅ Completed: SPOT5 benchmark (CNES 2001 satellite photography scheduling)
  • Dataset with problem instances
  • Standalone verifier for validation + scoring
• ✅ Completed: SatNet benchmark (2021 NASA/JPL DSN scheduling)
  • Dataset with problems and metadata
  • Standalone verifier for validation + scoring
  • Reference agentic baseline
• ✅ Refactored: Repository structure
  • Migrated to standalone benchmarks/ model (no inter-dependencies)
  • Deprecated abstraction layers (toolkits/, engines/) - using established
    libraries directly
• 🛠️ Environment: Migrated from uv → pixi (enables tudatpy integration)

Roadmap Context

Phase 1 (current): Legacy benchmarks

• ✅ spot5 (satellite photography scheduling)
• ✅ satnet (NASA/JPL DSN scheduling)
• 🔲 aeosbench (BUAA Earth observation)

Future phases: LEO constellations → Deep space → Rocket trajectories (if permits)

PR Purpose

This is a draft PR to track development progress. Main branch remains stable
snapshot.