Train: Optimised single_pass — charbonnier ramp + sky0.7 area-pooled + 50% full-noise (C2) 9ed518c3

Supersedes config 93 (“L1@10 + sky0.7 + 50% full-noise (C)”, run 2026-07-28_10-02-59_4779) after a conceptual review of commit 3cbdf26. Keeps the intent of (C) — supervised sky at 0.7 and 50% of examples at the zero-terminal-SNR endpoint — and fixes four issues that made it uninterpretable or actively harmful. 1. Prediction loss: the hard L2→L1 step at epoch 10 is replaced by an L2→Charbonnier ramp over epochs 10–14. A hard swap roughly doubles both primary losses in one epoch (E|d| = sqrt(2/pi)*sqrt(E[d^2])), shifting the primary/auxiliary balance and gradient clipping at the same instant; pure L1 also has non-vanishing gradient at the optimum, which leaves a dither floor — the opposite of the local-structure goal. Charbonnier is quadratic near zero and linear on large residuals. 2. LR schedule: constant_with_warmup → cosine_with_warmup. A constant LR is what makes the L1-family gradient floor bite; the template's own comment already recommends cosine here. 3. Latent weight pooling: new loss.depth_weight_pooling=area. With min pooling a latent cell mixing valid depth and sky collapses to the sky weight, so 0.7 attenuated exactly the thin structures against sky that this experiment targets. Area pooling keeps the all-or-nothing gate on genuinely missing pixels. 4. VKITTI2 valid_max_depth 300 → 650. VKITTI2 stores 16-bit centimetre depth, so its sky sentinel is 655.35 m; at 300 m the above_max fallback relabelled all real geometry between 300 m and the sentinel as sky and, at sky weight 0.7, actively taught it the 800 m override. Also enables depth.neutralize_invalid_depth on both training sources: under log_depth, clamping mapped 0/negative/-inf pixels onto the NEAR plane (1 m), and the VAE encoded that into neighbouring supervised latents despite zero loss weight. RDS retains valid_max_depth == sky_depth_override == 800, so its above_max sky mask may be empty. This is left unchanged deliberately because the raw RDS sky sentinel was not verified; training now warns at startup and logs realized per-source mask coverage each epoch. Check the rds `sky` fraction on epoch 0 before comparing sources. Requires the reviewed fixes on the training branch (depth_weight_pooling, charbonnier registry entry, neutralize_invalid_depth, all-breakpoint schedule validation, sky-inclusive smoothness region, and the depth_sky/nonsky prediction-loss split).