docs: 2026-09-30 arXiv sweep - citations, student-side closure, TODO.sota-2026

[ronaldtse]

Docs updates from the 2026-09-30 arXiv sweep (window Aug 26 - Sep 30: distillation, byte-level modeling, diacritization, optimizer literature).

No new competitor: zero text-only Arabic diacritization systems appeared on SadeedDiac-25 in the window; the field's Arabic-diacritization energy moved to the speech modality (KSAA-2026, not protocol-comparable). r7 remains best dedicated under our protocol.

Three commits:

1. Sweep citations (paper.adoc + RESULTS.md):
   • arXiv 2609.12303 (Meta/FAIR "Breaking the Token Ceiling") - distilled byte students reach a higher asymptote (+4% predicted) at 6x data efficiency; external scaling-law validation of our byte-student lineage.
   • arXiv 2609.37510 (MAESTRO) - teacher intervention in on-policy distillation injects off-policy load; always-on intervention is the worst point of the axis. The mechanistic account of our measured GKD negative (6.0036); the on-policy lever stays closed without a re-run.
   • arXiv 2609.10153 (YallaMorph, EMNLP 2026) - the concrete teacher-side data lever for the next teacher rung (run-009-yallamorph, in TODO.sota-2026/01).
2. Student-side lever family closed (RESULTS.md + paper.adoc caveat): complete residual ledger (8 closed levers), standing no-GPU-spend rule, and the honest single-seed caveat per arXiv 2608.27729 (per-seed sigma 2.8-48.7pp swallows sub-point deltas; our bootstrap resamples predictions, not seeds).
3. TODO.sota-2026/: five campaign spec files (r9-yallamorph teacher, citations record, closure rule, Stoicheia diffusion WO queue, RIDE direction probe).