Problem: Monolithic LLM generation of large structured outputs (N-cell grids) truncates, bloats per-cell with position, and cannot resume after a cutoff. Hypothesis: Locking every cell to a fixed word-bounded 4-tuple [ID|Name | Definition | 4 Steps | Novelty/Edge], emitting a one-line skeleton first and an in-band [CHECKPOINT k/K] marker at each stage boundary, holds per-cell cost at O(1) (~62 tokens), total at ~62xN, per-cell drift <3%, and allows deterministic resume at (stage k, cell j). Method: Independent A/B on 36-cell and 80-cell grids, real tokenizer counts. Claim scope: orchestration level only — O(N^2/k) attention, not O(n) attention. Proposed academic name: Schema-Constrained, Checkpointed State-Machine Generation (a.k.a. O(n) Segmented State-Machine Streaming / Segmented Multi-Turn Continuation with Explicit State Tracking). Forked from: pageman/from-kubota-s-cubic-theta-function-to-a-6-6-discovery-matrix-met on OpenVersions — Part 0 / Stage VI constructs VI.4 and VI.6, and the 6x6 Discovery Matrix run that first demonstrated the protocol.