Add adapter.pre_norm: optional norm ahead of the mlp_2layer projection

[amazloumi]

Summary

• AdapterConfig.pre_norm: a norm registry key, validated at config time against the registered norms; "" (default) builds nothing, so existing configs keep their state dict and init.
• MLP2LayerAdapter builds it as ln_q over the vision features and applies it ahead of proj1; reset_parameters re-initializes it through the norm's own reset_parameters(), which RMSNorm now provides. The other adapter types ignore the field.
• The registry reference lists adapter.pre_norm under the norm category.

Testing

• uv run ruff check kempnerforge/ tests/ passes
• uv run ruff format --check kempnerforge/ tests/ scripts/ passes (170 files)
• uv run pyright kempnerforge/ passes (0 errors)
• uv run pytest tests/unit/ -v --timeout=60 passes: 1905 passed, 3 skipped (33 new)
• If distributed code changed: uv run torchrun --nproc_per_node=4 -m pytest tests/distributed/ -v: 102 passed, 2 skipped on each of 4 ranks, same as main
• If training loop / parallelism / optimizers changed: uv run pytest tests/e2e/ --e2e -v: 27 passed, 4 failed, 1 skipped; the 4 (test_checkpoint_save_and_resume, test_moe_checkpoint_resume, test_pp_checkpoint_save_and_resume, test_sigterm_triggers_emergency_checkpoint) fail on main too
• uv run pytest tests/integration/ -v --timeout=120: 103 passed, same as main
• uv run pytest tests/smoke/ --smoke --data-path <pre-tokenized dir> --file-pattern 'shard_*.npy' --data-vocab-size 32000 -v: 19 passed
• uv run pytest examples/vlm/tests -q: 33 passed; uv run pytest examples/vlm/eval/tests/unit -q: 104 passed
• uv run sphinx-build -W --keep-going -b html docs docs/_build/html succeeds
• Diff coverage (--cov --cov-branch): 18/18 changed statements and 9/9 branch arcs in kempnerforge/
• Defaults unchanged vs main, seed 1234: init state dict, fp32 forward and all gradients byte-identical; a 4-GPU 20-step VLM run is loss-identical at every step
• pre_norm set, 4-GPU FSDP2: tests/distributed/test_vlm_fsdp.py passes with its adapter set to each registered norm; save → load restores outputs bit-exactly; resuming at step 10 matches the uninterrupted run through step 20
• Checkpoint mismatch: a no-norm checkpoint loaded into a pre_norm model raises, naming adapter.ln_q.weight; a pre_norm checkpoint loaded into a no-norm model loads and drops the norm, as it does for any parameter-adding field (e.g. model.qk_norm)

Closes #210

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

Files with missing lines	Coverage Δ
kempnerforge/config/adapter.py	100.00% <100.00%> (ø)
kempnerforge/model/adapter.py	97.53% <100.00%> (+0.09%)	⬆️
kempnerforge/model/norm.py	100.00% <100.00%> (ø)

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[Naeemkh]

I verified the compatibility claim rather than taking it on trust: with no norm configured the attribute never enters the module registry, so the state dict and init are unchanged for existing configs. All four builders already swallow unknown keyword arguments, so passing the new field to every adapter type is harmless. Adding the reset method to RMSNorm is also correctly scoped, since the only caller is the adapter re-init on a meta-device build.

One thing worth closing. Config validation checks that the value names a registered norm, but the adapter re-init additionally requires that norm to provide a reset method. Both registered norms do after this PR, so nothing is broken today, but the registry is meant to be extended and the next norm added without one would fail inside FSDP setup instead of at config time. Either guard the call or validate the contract alongside the registry lookup.

Minor: the norm is built with the default epsilon with no way to configure it, while model norms take theirs from config.