Finalize Run B training results and curriculum figure

[aadarwal]

Summary

• Replace the obsolete training lineages in Section V with the completed raw-base Run B protocol, shaped reward, fixed probe, and iterative deployment results.
• Replace the old gallery with one hash-locked L0 through L4 training figure built from the 150-step ledgers and six fixed 80-task probes.
• Update the abstract, Table III, the conclusion training summary, and the compiled paper asset to the validated 0.422 to 0.491 iterative result.
• Remove the gallery-only generator and raster assets.

Evidence and scope

• Run B probe JSON files are the source of record. W&B is not used as evidence.
• The post-L3 probe is retained exactly as measured and marked as affected by the archived harvest and evaluation contention window.
• No manuscript content outside the abstract, Section V, and the corresponding training summary in the conclusion was changed.

Validation

• Figure asset preparation and hash validation passed against runs-v2/runB-base-rl.
• Ruff passed on both figure scripts.
• Targeted RL protocol tests: 16 passed, 4 skipped.
• Tectonic compiled the 16-page paper with no overfull boxes, undefined citations, or undefined references.
• The abstract and pages 12 through 14 were visually inspected, and assets/pixcell-paper.pdf matches the compiled PDF byte for byte.