We present an independent end-to-end reanalysis of the Atacama Cosmology Telescope Data Release 6 (ACT DR6) daytime and nighttime sky maps across the PA4 (150, 220 GHz), PA5 (90, 150 GHz), and PA6 (90, 150 GHz) detector arrays, spanning angular scales 600 ≤ ℓ ≤ 2500 and combining cross-split power-spectrum reconstruction, map-level time-split null tests, Spin-2 polarization diagnostics, beam-systematic modeling, and GPU-accelerated Bayesian cosmological inference. Evaluating mode-decoupled pseudo-C_ℓ estimators on the Needlet Internal Linear Combination (NILC) analysis footprint (f_sky = 0.312) with point-source-subtracted maps, nominal instrument beams, and exact SACC bandpower window functions, we reproduce the official ACT DR6 nighttime auto-frequency bandpowers to within 0.1%–1.4% (|ΔD_ℓ / D_ℓ^off| ≤ 0.006 across 800 ≤ ℓ ≤ 2000 for PA5 90 GHz, PA5 150 GHz, and PA6 150 GHz) and all six cross-frequency channels to within 0.2%–2.0% once co-temporal split pairs (i = j) are excluded to eliminate correlated atmospheric noise. Map-level four-way time-split null tests 〈(s0 - s1) × (s2 - s3)〉 and 〈(s0 - s2) × (s1 - s3)〉 pass across the full multipole range with probabilities to exceed PTE = 0.45 and 0.30, respectively. For the daytime dataset, we derive unbiased relative gain calibration factors c_day = 〈D_ℓ^{n0 × n1} / D_ℓ^{d × n}〉 from day × night cross-spectra across all six array-frequency bands, obtaining c_day = 1.030 ± 0.003 (PA5 90 GHz) and c_day = 0.992 ± 0.003 (PA5 150 GHz) on the NILC footprint, along with unit split-gain invariants r_split confirming phase coherence between the daytime and nighttime sky maps. Direct Spin-2 spherical harmonic analysis of the daytime Stokes (Q, U) splits reveals parity-conserving and parity-violating residuals (|D_ℓ^BB| > |D_ℓ^EE| ~ 10^2–10^3 μK^2, |D_ℓ^EB| ≠ 0) two to three orders of magnitude above the cosmological E-mode signal and uncorrelated with nighttime polarization, whereas nighttime temperature-to-polarization leakage remains below 0.036%, establishing from first principles why daytime cosmological analyses must be restricted to temperature. We further investigate a 2σ high-ℓ deficit in PA6 90 GHz nighttime auto-spectra, showing that a 1% linear beam-transfer correction above ℓ = 1000 improves the full-covariance goodness of fit from χ^2/ndf = 2.066 (PTE = 5.3 × 10^-4) to χ^2/ndf = 0.675 (PTE = 0.910), and demonstrate via a six-pair split covariance analysis and 14-mode beam principal-component projection of PA5 90 GHz that a localized -6.7% residual at ℓ = 2375.5 originates from map-level compact-source subtraction rather than smooth beam modes. Incorporating our calibrated daytime and nighttime bandpowers into a differentiable neural-emulated ΛCDM likelihood, we find that daytime temperature data tighten TT-only parameter posteriors by 11.3% in H_0 and 12.1% in n_s, while contributing a ~1.0% tightening in the full TT + TE + EE likelihood where nighttime polarization dominates the Fisher weight. Marginalizing over the 14 PA5 90 GHz beam uncertainty modes shifts n_s by +0.83σ (0.9689 ± 0.0070 → 0.9765 ± 0.0092), highlighting the coupling between high-ℓ 90 GHz beam calibration and the scalar spectral index.