Quantum-Inspired Thermal Equilibrium: Integration of Collective Dynamics into Real-Time Stabilization of Magnetically Confined Fusion Plasmas — Extended Edition: Burn Control and Unstable Equilibria across TORAX, FreeGSNKE and FUSE, with Hardware-in-the-Loop on Four Microcontroller ISAs

Abstract

2026-09-25 — Second-transport-code extension of the deposit: the alpha point exists in FUSE. Chapters 35–37 held a thermally unstable burn point that the canonical 0D model places at T* = 3.59 keV; this chapter locates it in a second, independent 1D transport code — FUSE.jl v1.2.1 (General Atomics / ProjectTorreyPines: time-dependent transport with neural-network TGLF turbulence, EPED pedestal, current diffusion, sawteeth) — on ITER at 15 MA, over 180 stationary solves and 41 time-dependent runs. ITER burns: with the ITER-trained TGLF-NN model sat2_em_iter4_azf-1, 31 MW of auxiliary heating give 320 MW of fusion power, Q = 10.3, alpha heating 2.1 × the external heating, and the plasma stays ignited with the heating switched off (195–220 MW). THE UNSTABLE POINT: at zero heating a hot ignited and a cold state coexist at identical actuators, separated by a sharp threshold — a 45 % core cooling recovers to the burn, 50 % collapses to 2 MW, identically at half the time step; the trajectories on either side separate exponentially with γ = 0.19–0.27 s⁻¹ (e-folding 3.8–5.3 s), the signature of an unstable equilibrium, bracketed at ⟨Te⟩ ≈ 4.2–4.4 keV and 58–71 MW of fusion power — a far deeper physics model finds the threshold at a temperature similar to the 0D prediction. EARLY WARNING AND MEMORY: on the approach to the end of the burning branch the recovery time after a core kick rises from 2.4 s to 10.7 s (critical slowing down, readable by a controller); at identical 10.4 MW heating the plasma either burns at 23 MW or stays cold at 1.1 MW, depending on whether the heating was lost for 10 s; and heating moves the threshold — the same −50 % cooling that quenches the unheated burn is caught with 10.4 MW. QITE INSIDE FUSE: the engine library built on Linux from the exact source of the study DLL is bit-identical to it (engine, Q16 and Q8 paths; a 30 000-decision closed loop with identical SHA-256), and the unchanged five-agent chapter-35 fleet runs in the FUSE time loop with 300/300 decisions (7 500 values) bit-identical between Windows and Linux — the platform for holding this point with the collective. Declared limitations, stated in the chapter: the burning branch appears with the sat2/sat3 ITER-trained TGLF-NN models (the case default model stays heating-driven, Q = 1.5 at full heating); the EPED pedestal is imposed throughout (no H–L back-transition); no helium-ash build-up. Chapter 39 of the master document (6 pages, 3 figures, A4) is inline-embedded in Bernd25.09.2026.pdf; every reported number is recomputed from the shipped CSVs by the included analyze_ch39.py, and the scan runner and the Julia environment are shipped for reproduction. The aggregate deposit now contains 29 520 reproducible episodes across 105 (study × policy) groups. All measurement data and figures released under CC-BY-4.0 / CC0; only the QITE Core SDK itself remains proprietary (DPMA Aktenzeichen 10 2025 003 906.9 and 10 2025 004 891.2). Full technical specification in chapter 39. 2026-09-24 — Second-instability extension of the deposit: the collective leaves the burn timescale by three orders of magnitude. Chapters 35–37 studied the thermally unstable burn point (e-folding time ~30 s); this chapter holds a MAST-U-like plasma against its vertical displacement instability (growth rate 91.8 s⁻¹, e-folding time 10.9 ms) in FreeGSNKE 3.1.0 (UKAEA/STFC), a time-dependent free-boundary equilibrium code with active coils, passive structures and eddy currents — a second, independent simulator family next to TORAX. Actuator: the P6 coil pair in anti-series, 20 V limit; disturbance: a 40 V / 1 ms kick (~7.5 mm); loss: |Z| > 20 mm; uncontrolled, the plasma is lost within 5 ms. The collective (Q16, Q8 and FP32 arithmetic) is benchmarked over 14 283 closed-loop runs against FreeGSNKE's own tuned vertical controller (PD on Z·Ip) and a proportional controller, at decision rates from 54 down to 1 decision per e-folding time, 20 seeds each. BENCHMARK: with lead evidence (the predicted position, using the PD's own 2.5 ms lead time) the Q16 collective matches the tuned PD on the key number — all seeds held down to 3.03 decisions per e-folding time (one decision every 3.6 ms), none at 2.27 (cliff 2.73 vs 2.74) — on all four members of a plant family spanning tau = 3.2–19.1 ms (2.48 / 2.73 / 2.75 / 3.02 vs 2.29 / 2.74 / 2.88 / 3.03); the FP32 engine with lead evidence has the best steady-state precision of all controllers (0.063 mm vs 0.073 mm for the PD) at the lowest voltage and zero saturation; with position evidence the collective catches the kick with the smallest excursion (3.5 mm vs 4.2 mm PD, 8.2 mm proportional). The decision-rate coordinate of chapters 36–37 transfers: for every controller the cliff stays constant in decisions per e-folding time across a six-fold range of tau (spread 1.22–1.32). LIMITS: the FP32 engine holds the plasma for 10 s (918 e-folding times); an 80 V kick (~14 mm) is caught by all controllers; at 100 V the FP32 engine still holds 2/5 where Q16 and PD hold 0/5. XHARD-VERTICAL (protocol and expectations hashed before the campaign): 1 920 runs — six controllers × four plants × eight severities × ten seeds, 50 e-folding times each, every controller on its reference tuning with no retuning, under a paired fault cascade of ELM kicks with pressure crashes, supply dropouts, supply-efficiency drift, sensor noise, bias, frozen readings and latency: the two lead variants survive as often as FreeGSNKE's PD (0.497 and 0.494 vs 0.503; exact McNemar p = 0.75 and 0.25 on identical cascades), and on the slowest plant the FP32 lead variant has the highest survival of all. HARDWARE: the STM32F411 (ARM Cortex-M4) reproduces the host computation bit for bit (40/40 runs) and holds the plasma at one decision every 3.6 ms; the ESP32-C3 (RISC-V) holds it for 110 ms at 10.9 decisions per e-folding time. Numerics: fast linear mode vs full Grad–Shafranov reconstruction 33/33 identical outcomes; halving the solver step shifts the cliff by 0.006; a neutral harness shipped with the data reproduces the PD, proportional and XHARD rows bit for bit. A decision-rate prediction written and hashed before the campaign is reported in full in the chapter. Chapter 38 of the master document (6 pages, 3 figures, A4) is inline-embedded in Bernd25.09.2026.pdf; every reported number is recomputed from the shipped CSVs by the included analyze_ch38.py. The aggregate deposit now contains 29 479 reproducible episodes across 100 (study × policy) groups. All measurement data and figures released under CC-BY-4.0 / CC0; only the QITE Core SDK itself remains proprietary (DPMA Aktenzeichen 10 2025 003 906.9 and 10 2025 004 891.2). Full technical specification in chapter 38. 2026-08-02 — Validation extension of the deposit: the law holds, and the unstable branch becomes a road. Chapter 36 measured a decision-rate law on the canonical 0D model (existence boundary ~4.3 decisions per e-folding time, migrating logarithmically with sensor noise); this chapter takes it to the full TORAX 1.4 transport plant and asks it to predict something it has never seen. Five decision intervals (0.5–8 s) under the XHARD severity-5 cascade with the disturbance dose normalised per second (per-tick draw probabilities rescaled as p_tick = 1−(1−p_1s)^dt, random walks by sqrt(dt) — without this, slower ticks would silently reduce the dose), ten seeds each, the LX6 chip computing every decision: survival falls from 9/10 at 3.7 decisions per confinement time to 1/10 at 1.85 and 0/10 below — the transport plant's cliff sits where the 0D boundary said it would, across a change of model, physics and timescale proxy. At the crossing, a first for the series: one plasma ignited cleanly, burned 500 s and was lost from flat-top (t = 630 s) — through chapters 31–36 every plasma that completed its rampup completed its pulse; below the rate boundary that rule ends, exactly as the law requires. STRAND TWO: 20/20 commanded full burn cycles on the chip — ignite from the cold branch, park ON the provably unstable point of chapter 35, full burn, controlled shutdown, re-ignite, return — 120/120 waypoints reached (median settle at the 90 s ramp end, worst +105 s), zero disruptions, reproduced 20/20 by the simulation twin: burn access and controlled ramp-down executed as routine through the region every machine fears. STRAND THREE, from existing grids with no new experiments: the boundary belongs to the task — the cleaned classical controller (PI, derivative term removed) meets the collective's wall within 9 % at the one point where both are measurable (tau/dt 7.6 vs 8.3 at 16 % noise); the deviants deviate for named mechanisms (the full PID's derivative term amplifies sensor noise into a binary wall with no frontier at all; memory-less bang-bang pays explosive rates, 6.1 to 17.8 per noise doubling), and a second operating point identifies actuator authority per e-folding time as the third dimensionless group of the law. Chapter 37 of the master document (5 pages, 3 figures, A4) is inline-embedded in Bernd02.08.2026.pdf; every reported number is recomputed from the shipped CSVs by the included analyze_frontier.py. The aggregate deposit now contains 15 196 reproducible episodes across 53 (study × policy) groups. All measurement data and figures released under CC-BY-4.0 / CC0; only the QITE Core SDK itself remains proprietary (DPMA Aktenzeichen 10 2025 003 906.9 and 10 2025 004 891.2). Full technical specification in chapter 37. 2026-08-01 — Scaling-law extension of the deposit: the price of a decision. Chapter 35 established that the collective balances the thermally unstable burn equilibrium (T* = 3.59 keV, e-folding time tau = 30.5 s); this chapter measures the quantitative law behind that result, entirely on the fourth-ISA hardware (classic ESP32, Xtensa LX6, Q16 collapse on chip): 1 130 closed-loop episodes over four single-factor axes around the chapter-35 reference (decision interval 0.5–24 s, sensor noise 1–32 %,