Measuring the MOND Acceleration Scale at High Redshift: The Baryon-Calibration Wall

v1.3: corrects the Gaia DR4 paragraph. Version 1.2 called Arm A's floor (1.1614, alt 1.1917; nu_RAR kernel) and its pair counts those of 'the bare law'. With the framework's own P2 kernel the merged law gives 1.063-1.102 canonical (1.079-1.127 alt), 2.3-3.7 sigma from ownership at N = 30,000. Found by an independent referee (CFG265) of a companion note. The a0(z) content is unchanged. Version 1.2: wording corrections after an independent referee (CFG241; 23 findings, 0 critical, 2 major); no number changes apart from one mislabelled threshold (2 SD = 0.168, with SD = 0.084), whose conclusion is unchanged. The programme ties the MOND acceleration scale to the vacuum, a0 = kappa c sqrt(G rho_Lambda), with kappa = 1/2 fitted, not derived. Its distinctive high-redshift prediction is a flat a0(z) (distinctive against LambdaCDM and the H(z) rival; standard MOND shares it); the rival is a0 proportional to H(z) (about x2.2 at z~1.4, x8 at z~5); LambdaCDM has no a0 and enters only through an effective-a0 proxy. We collect the programme's committed high-redshift a0 lanes and state one result plainly: in every lane the limit that survives an unlimited sample is the absolute calibration of the baryon (gas and stellar) masses; at present sizes the KiDS split, the class-S sources of the compilation and the seven class-M galaxies are also limited by statistical power or by the Newtonian regime (CFG255, CFG227, CFG229). The lever. Where the discs sit (per-point median g_bar/a0 = 1.1 to 4.4) a 0.05 dex error in the mass discrepancy is a factor 1.5 to 1.8 in a0; the implied a0 moves by -2.5 to -4.8 dex per dex of baryon mass (a post hoc finite-difference estimate at a +-0.03 dex step; -2.5 to -6.4 at +-0.01 dex). In the deep regime an amplitude test measures only the product of the calibration and a0. Lean theorems (ChainCert, 331 theorems) show that a0 remains identifiable in principle for the exact kernel, but the estimate is ill-conditioned: with the calibration free, and for local kernel slopes in [0, 1/2] with independent Gaussian errors, sigma(log a0) >= 3 sigma/sqrt(N); a log-uniform sample of 20 galaxies at 0.1 dex must reach g_bar/a0 of about 8 and include deep points to measure a0 to 0.1 dex. The samples. No sample separates the laws: RC100 is gas-route-limited, CRISTAL is route-dependent, and the z = 2-5 compilation, the ALMA [CII] rotators and seven class-M galaxies are NOT POSSIBLE at pre-flight. The gas calibration at z~2.2 is a prescription bracket (~0.2-0.7 dex), and no public dynamics-independent anchor reaches 0.10 dex: the best route, an absorption-line alpha_[CI](Z) relation, does not reproduce its stated 0.2 dex scatter from our transcription of its Table 1 (0.64 dex about the quoted relation; how the authors define that scatter, and any erratum, were not checked; an apparent cyclic shift in four rows is a hypothesis not put to the authors). A within-survey lens-redshift split of the KiDS-1000 lensing RAR, which would cancel a common-mode calibration, is NOT POSSIBLE (power 0.14 against 9) and non-discriminating by its own control; even with unlimited statistics a stellar-mass drift delta would mimic an a0 drift by delta/2. Bottom line. A decisive a0(z) test needs baryon errors and band at or below 0.10 dex. With the public data on hand the question is open; nothing here is a detection either way, and nothing here says the data favour the framework, the rival or LambdaCDM. The near-term test of the programme is a separate z = 0 one, which its current candidate law can only survive or fail: Gaia DR4 wide binaries (2 December 2026). Every number carried by an audit row is re-read from committed files by PAPER38_audit.py (331 of 331); numbers quoted only in running text were checked by the referee, not by the script. AI-assisted research programme; not peer reviewed; nothing is claimed closed.