[Proposal] Honor cfg.init_weights=False in boot_native and add bridge.init_weights()

[jlarson4]

Proposal

Base on dev-4.x. Gate the initialize_native_model(...) call in boot on cfg.init_weights (transformer_lens/model_bridge/sources/native/__init__.py:56-62), and add a public TransformerBridge.init_weights() re-init method.

Motivation

Deprecating HookedTransformer moves train-from-scratch to boot_native, but the bridge ignores cfg.init_weights=False (randomizing weights anyway) and has no re-init entry point, breaking the load-into-empty-shell workflow.

Pitch

• Gate initialize_native_model(...) on cfg.init_weights, keeping fork_rng for the seeded branch when init runs.
• Add TransformerBridge.init_weights() re-running initialize_native_model in place, honoring cfg.seed.
• Raise a clear error when init_weights() is called on a non-native (HF-loaded) bridge.

Acceptance:

• boot_native(cfg) with init_weights=False leaves default nn.Linear/nn.Embedding reset state
• bridge.init_weights() re-inits a native bridge in place, honoring cfg.seed
• init_weights() on an HF-loaded bridge raises an actionable error
• New unit tests in tests/unit/model_bridge/test_boot_native.py (skip + re-init paths). The file already exists, it just needs to be extended
• make unit-test and uv run mypy . pass

Mirror: HT gate transformer_lens/HookedTransformer.py:254-255; HT init_weights at HookedTransformer.py:1500.

Checklist

• I have checked that there is no similar issue in the repo (required)

[emerardd]

Hi @jlarson4, I'd like to take this, based on dev-4.x.

I confirmed that the native boot path currently calls initialize_native_model(...) unconditionally, so cfg.init_weights=False is not honored.

My planned scope is:

• gate the custom native initialization while leaving PyTorch's default module initialization intact;
• add TransformerBridge.init_weights() for native bridges, preserving the existing scoped-generator behavior so cfg.seed is reproducible without perturbing global RNG;
• raise an actionable error for non-native bridges;
• add regression coverage for the skipped-init, seeded re-init, non-native error, and RNG behavior.

I'll keep this independent of #1568 and won't change the native initialization math.

[jlarson4]

@emerardd sorry looks like @happykawayigt has already put in a PR for this!