Coverage figure: sixteen cells, all learn - Gaussian square (E40), fpo-policy · FPO · square (E41), square plots success

[tactino]

Stacked on #96 (E41): merge that first, and this diff shrinks to the figure files and the READMEs.

The coverage figure is now sixteen cells, and all sixteen learn.

• gaussian-ppo-square is new: DPPO's Gaussian MLP fine-tuned from its released checkpoint with DPPO's PPO settings (E40).
• fpo-fpo-square learns: E39 resumed to FPO++'s 8M steps by E41, two of three seeds past the bar. Its curve is E39's 100 iterations followed by E41's 67 (new resumed_from in cells.json), and its clip is E41's seed 1.
• Square plots the success rate. Its runs do not share a reward: DPPO's settings count every step after success, and FPO++'s end the episode on it. On one return axis, FPO++'s 0-1 returns drew as a flat line under the DPPO cells' 50-160. Every square run logs rollout/success.
• The pi0.5 cells get E36's numbers. pi0-policy · DPPO's note now says 20 of 50 is below the untrained 31 ("barely moved it" was wrong).
• The READMEs say sixteen cells, and that pi0.5 on LIBERO now has its own page on the site rather than sitting under the grid.

The site PR that publishes this follows.