Honor native bridge weight initialization config

[happykawayigt]

Description

Honors cfg.init_weights=False in TransformerBridge.boot_native() and adds a public TransformerBridge.init_weights() method for deterministic, in-place reinitialization of TL-native bridges.

Bridge construction replaces native submodules with generalized component wrappers, so the native initializer now unwraps those components before resetting the original parameters. Calling init_weights() on a non-native bridge raises an actionable error pointing to boot_native().

Fixes #1566

Type of change

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• This change requires a documentation update

Validation

• pytest tests/unit/model_bridge/test_boot_native.py tests/unit/model_bridge/test_build_bridge_from_module.py tests/unit/model_bridge/test_native_features.py -q: 78 passed
• mypy .: Success, no issues found in 421 source files
• Full pytest tests/unit -m "not slow" -q: 5142 passed; the only two failures were existing UTF-8 test files read with the Windows GBK locale, and both pass with PYTHONUTF8=1
• Black, isort, pycln, and git diff --check

Development note: AI-assisted implementation; I manually traced bridge component replacement, reviewed the patch, and ran the focused and repository-wide validation above.

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

[emerardd]

Thanks for working on this. I tested the re-init path locally and found two edge cases worth checking:

UnembeddingBridge adds a zero bias to the native bias-free head during bridge construction. The current test zeroes all parameters before re-init, so it cannot detect that this bias is not reinitialized. Filling parameters with a non-zero value such as 42 shows that head.bias remains stale unless initialize_native_model() explicitly zeros it.

With init_weights=False, the current seeded branch still enters fork_rng, but cfg.seed is never applied to NativeModel construction. As a result, different non-null seed values produce identical default parameters and only affect whether global RNG state is restored. The issue wording may imply that the seeded fork_rng should only be used when custom initialization runs.

I also exercised re-init across plain/RMS/no-norm, gated MLP, attention-only, rotary, separate-output-vocab, and all five initialization modes; those paths otherwise worked correctly.

[jlarson4]

Thanks for closing this out @happykawayigt! Great resolution. Just one small note:

HookedTransformer.init_weights() now has a bridge equivalent, but the migration guide never mentions init_weights. There is a table for API table equivalents in Migrating to v3 (migrating_to_v3.md:223). Could you add that row for cfg.init_weights / bridge.init_weights()?

[happykawayigt]

Added in 3089115: the migration table now maps cfg.init_weights to bridge.init_weights() and notes that boot_native() honors the config flag during construction. The updated checks are green. Thanks for catching this.