Number training rewards and condense research contracts

[aadarwal]

Summary

• Number the two reward branches as Eqs. (4a) and (4b).
• Replace the long Research Contracts section with a concise conclusion and four research directions, including why each direction follows from the paper.
• Keep mutable contract criteria in the versioned registry and direct readers to the project page for live claim state and claim controls.
• Refresh assets/pixcell-paper.pdf from the compiled manuscript.

Validation

• Compiled arxiv/main.tex with Tectonic.
• Visually inspected the affected equation and conclusion pages in the rendered PDF.
• Ran uv run pytest tests/test_research_contracts.py -q: 31 passed, 20 xfailed.
• Checked the live contract registry and claim flow on the PixCell project page.