We consider an SU(3)dark gauge theory with three flavours of heavy dark quarks of mass M_Q ∼ 10^13—10^15 GeV, neutral under all Standard Model gauge groups. The theory is assumed to undergo a first-order confinement transition at T_c^dark ≈ 10^15 GeV—a phenomenological input, not a first-principles result, carrying a ±50% theoretical uncertainty from the bag-model approximation in analogy with the Columbia-plot first-order region of QCD at heavy quark masses. We compute the resulting stochastic gravitational-wave (GW) background, finding that acoustic sound waves dominate by two orders of magnitude over bubble collisions and magnetohydrodynamic turbulence after the finite sound-wave lifetime suppression Υ ≈ 0.05 is applied. At the benchmark parameters—latent heat fraction α_PT = 0.1, dimensionless inverse transition duration β/H* = 100, and bubble wall velocity v_w = 0.8—the composite peak falls at f_* ≈ 1.65 mHz with Ω_GW^peak h^2 ≈ 10^(-8), and the sub-peak slope Ω_GW ∝ f^3 follows from causality independently of the phase-transition parameters. Under the assumption that the signal is present and that LISA achieves its projected sensitivity, four years of observation would yield a signal-to-noise ratio SNR ≈ 8.5, with a ±18% systematic uncertainty from the sound-wave lifetime calibration already absorbed into the quoted value. A Fisher analysis suggests that the peak frequency and amplitude could be recovered to approximately 15% and 25% respectively, which would constrain the dark confinement scale Λ_dark to within a factor of two—assuming the one-loop dimensional-transmutation relation holds. We note that pulsar-timing arrays, CMB spectral distortions, and direct-detection searches test the same parameter space from different directions; whether they can be combined into a joint exclusion is a question we leave to future work.