Across 300 sampled payloads from today, `phi_c_estimate` was pinned at the 3.2 scan ceiling with tiny bracket widths, usually about 6.9e-06 to 9.8e-06 at 2048 and only about 0.0017 to 0.0067 in the sampled 128-site cases. The mode was strongly bridge-loaded but never rescuing: `last_mean_bridge_ratio` sat around 9.7 to 10.5 in deep runs, `last_mean_canal_mass` stayed high around 0.545 to 0.556, and hot-refuge balance often hovered just under target around 0.44 to 0.47, occasionally brushing about 0.48. But `last_mean_max_real_part` remained negative in every sampled payload, typically about -0.50 at 2048 and around -0.85 to -0.89 in the smaller runs, so the support flag stayed 0.0 even when the geometry looked nearly right.