[jlarson4]

Description

• Updating our adapter and contributor documentation to reflect lessons learned from new bug fixes and to reflect changes in how we approach things due to those fixes
• Added additional tests to cover gaps that were not yet being tested
• Updated ActivationCache to cover edge cases and improve DRY state
• Updated SVDInterpreter with a lazy load to reduce base transformer_lens` load
• Refactored bridge_core.py to be able to determine _registered_block_count with a method, to reduce the repeated calculations around that value
• Adjusted the stop_layer system. Now processed as part of the bridge, where the indexes are already known, so they do not need to be reconstructed by the block components
• Created an _eps helper for normalization
• Fixed incorrect configurations for Gemma3n and MuseGlimmer
• Added validation to patching
• Removed the static _HYBRID_VARIANT_NAMES set and replaced with dynamic hybrid detection
• Improved warning/error accuracy for DLA & other analysis tools
• Refactored statistics to use sub-functions instead of mulitlayer if and variable declarations

Changes in behavior

• Right-padded string-batch loss now includes the transition into a real trailing pad-id token, those loss values change and now equal the mean of the per-prompt token losses.
• input_to_embed on a padded string batch returns the tokenizer's mask for either padding side. Before, a right-padding tokenizer got None and only a left-padding tokenizer got a mask (rebuilt from the ids).
• input_to_embed(..., padding_side="right") on a left-padding tokenizer used to ignore the argument when building the mask, so it returned a left-side mask read off right-padded ids. It now returns the mask for the side used.
• Statistics on GQA-padded cells become NaN; to_dataframe() row counts change on GQA results.
• DLA no longer raises NotImplementedError for hybrids, it succeeds or raises ValueError.
• Gemma 3n SVD w_in outputs change for every checkpoint
• Muse Glimmer gains blocks.{i}.ln1/ln1_post/ln2/ln2_post.hook_scale/hook_normalized, and its post-norm head/neuron decompositions now return output.
• ModernBERT-decoder lenses are refused at validate_model/fit.
• run_with_cache(stop_at_layer=k) raises NotImplementedError on bridges without a blocks stack.
• start_at_layer resumes correctly on Bloom and opaque/SSM stacks, and raises NotImplementedError on AltUp blocks.
• ±inf effects give NaN p-values (was p = 0 or 1/(n_perm+1)). DLA unit="component" refuses an incomplete decomposition before the forward (was after a full forward).

Type of change

• Bug fix (non-breaking change which fixes an issue)
• Breaking change
• This change requires a documentation update

Checklist:

• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes
• I have not rewritten tests relating to key interfaces which would affect backward compatibility