docs: Phase 9 cross-domain forgetting matrix (dsp)

docs: Phase 7 CUDA bench results (dsp)

docs: add upstream Qwen3-4B-Instruct-2507 official evals

docs: add base vs tuned bench comparison

docs: standardize Phase 6 bench table (mascarade column)

docs: replace Phase 6 Gemma bench with real Qwen3-4B held-out eval (dsp)

docs: add Benchmark / Training metrics section

Training in progress, step 100