docs(papers): campaign close — frontier rows, residual closure, static-int8 artifact

[ronaldtse]

What

Closes the papers against the completed campaign:

paper.adoc

• Frontier table: three post-2.1 rows (Engram 4.67 [2.03, 2.47] · headwise Muon 4.82 [2.15, 2.55] · Sinkhorn 4.55 [1.84, 2.23]), teacher-delta CIs from each run's final_eval
• Residual paragraph: records the optimizer-recipe arms transferring negatively at 300M — head-wise Muon separated-worse (+0.2267pp [0.016, 0.411], paired between-students), Sinkhorn indistinguishable from AdamW — closing the optimizer-artifact hypothesis alongside the corpus hypotheses
• Discussion: the on-policy "open lever" is now its measured negative (6.0036); the path restated as teacher-side data (r5→r6→r7: −0.10pp aux-task, −0.29pp domain mix), next experiment an r8 teacher

paper-c.adoc

• Artifacts section records the shipped static-int8 composition: ara-diac-small-int8static-2.1, 491MiB, cer_delta 0.1038pp over 2,480 samples, zero confident-position flips