Add large-sparse-shape stress gate for SLat students

Advance v7 campaign to accepted 10-to-4 gates and 4-to-2 training

Record stable 10-to-4 training and SS held-out gate

Add gated progressive stage planner

Add fail-closed v7 quality gate checker

Add v7 progressive distillation quality gates

Record 10-to-4 gates and start 4-to-2 stage

Make held-out evaluation aware of progressive adapter chains

Add size-aware held-out evaluation for sparse students

Use FP32 master LoRA weights for stable L4 distillation

Expose stable BF16 and LR controls in L4 trainer

Harden v7 distillation with BF16 and non-finite guards

Add VRAM-aware SLat curriculum controls

Fix AniGen SparseTensor symmetry transforms

Advance v7 campaign to short 10-to-4 training

Record MoE, multiview and sparse24 recovery L4 validations

Track first v7 real companion training campaign

Add standalone v7 TensorRT student builder

Split v7 FP8 quantization from TensorRT build

Add held-out teacher/student evaluation for v7 adapters

Add deterministic train/eval split to v7 cache campaigns

Add batched AniGen teacher-cache builder for v7 campaigns

Add batched FLUX.2 v7 companion reference campaign generator

Place injected Geometry MoE experts on AniGen device and dtype

Complete ModelOpt calibration reader API

Make Geometry MoE sparse-aware for AniGen SLat FFNs

Add real SLat 2:4 recovery L4 smoke

Add real SLat selective FP8 QAT L4 smoke

Add real SLat Geometry-MoE L4 smoke

Auto-install ModelOpt ONNX extras for calibrated FP8 builds

Add exact 2:4 shared-teacher recovery with final re-prune

Add selective FP8 QAT shared-teacher recovery

Add real calibration capture smoke for chained one-step students

Select SDPA before AniGen import for SS ONNX export

Add real four-view AniGen teacher cache builder

Isolate AniGen cache teardown with optional child hard-exit

Auto-install onnxscript for SS opset23 student export

Keep MoE router probabilities differentiable for load balancing loss

Add Geometry-Inductive MoE shared-teacher training loop

Make Geometry MoE trainable with shared-base teacher mode

Run symmetry guidance every 8 distillation steps on L4

Backprop rig-aware flow symmetry prior during v7 distillation

Integrate flow equivariance symmetry loss for v7 distillation

Add chained one-step student ONNX export smoke

Add reproducible chained 2-to-1 L4 smoke worker

Allow isolated smoke namespaces for v7 student artifacts

Mark SS and SLat few-step backward validated on L4

Add calibrated FP8 and 2:4 TensorRT build pipeline for v7 students

Add real SS/SLat calibration capture for v7 FP8

Add chained v7 student ONNX export for SS and SLat