autograd.py: replace _factored_weight_norm with delta-based norm at merge time

Fix dtype mismatch: promote-then-downcast for mixed bf16 base + fp32 adapters

README: correct factored-norm PR reference — smellslikeml/peft#8, not upstream

README: sharpen Downstream reach per Appendix G Table 15

README: rehost composite on gist + name PEFT's downstream reach

README: embed algorithmic-analysis composite (dense vs. factored)

README: license status honest — pending upstream authorization

README: link the Hub-consumed A100 validation Colab

README: rewrite for external readers — drop internal Stage A/B/C narrative

metadata.json: backend=cuda with arch list

metadata.json: kebab-case name, simplified id

Add build/torch-universal/metadata.json for get_kernel() consumption

Add kernels library_name tag to README frontmatter

Reshape for kernels-community pure-Triton publish

Fix ruff lint + formatting on Stage C files

Drop strided-direct tests that exercise unsupported input orientation

PR #16 refinement — strided-test broadcast fix + wrapper allocation reduction

DoRA-factored kernels-library package — Stage C PEFT-loader wiring + transpose-overhead fix

DoRA-factored kernels-library package — Stage B Triton fast-path (#15)

DoRA-factored kernels-library package — Stage A scaffold (#14)