Final deprecation prep

[jlarson4]

Description

Hook API & bridge correctness

Gated-hook centralization – check_hooks_to_add is the real gate and keys on the hook point's canonical identity, so aliased overrides that genuinely fire (BERT's hook_mlp_in) are no longer refused, filter sweeps warn-and-skip consistently across add_hook/hooks()/run_with_hooks, and the default caching sweep composes through hooks() without raising.
W_pos / W_E_pos bridge accessors – last WS-3 item; refuses T5's relative-bias table, slices OPT/BART's +2 positional rows to match the legacy converters.
hook_pattern write-capable on native bridges – the hook runs inside NativeAttention before the value matmul, so pattern edits genuinely re-weight the output instead of silently no-oping.
Deprecation warnings fire exactly once per entry point – from_pretrained suppresses the init warning, from_pretrained_no_processing warns at its own stack level so the default filter shows it, and the two classes that never warned on bare construction now do.
Legacy-model support on the bridge

boot_tl_legacy – loads legacy TransformerLens-format repos (NeelNanda/ArthurConmy/Baidicoot: TL config.json + .pth, file-based checkpoints, old-SoLU and minGPT layout converters ported in); SoLU-1L is bit-exact against the frozen goldens.
Native solu_ln – the SoLU activation plus the load-bearing mid-MLP LayerNorm, without which those checkpoints compute the wrong function.
Checkpoint plumbing – NeelNanda file-based labels restored to the registry (WS-6 regression), cfg.checkpoint_index/value stamped by both boots, negative indices accepted for legacy parity.
Wav2Vec2 adapter – reuses the HuBERT structure, routes the ForPreTraining checkpoints to their encoder, refuses direct pretraining-head wraps, and fixes encoder_output's block order for stable-LN encoders.
Tests re-anchored off live HookedTransformer

incl_bwd gradients – anchored on a new goldens gradients group (dataset revision bumped) plus an exact autograd self-consistency test.
Checkpoint-converter tests – run from frozen in-repo fixtures instead of building nine HookedTransformers.
BERT weight surface – all nineteen accessors certified on the bridge against raw HF parameters, succeeding the HookedEncoder-only suite.
Enc-dec stacking parity – oracle derived from raw HF T5 weights instead of a live HookedEncoderDecoder.
Demos

All ten Hooked-touching demos migrated – Main_Demo's checkpoint loop runs its original SoLU model via boot_tl_legacy; GPT_OSS's bespoke 60%-loader collapses into one boot call (its hand-built config was unfaithful — the rewrite fixes real divergence); Colab_Compatibility is reframed as the Legacy Compatibility Ledger tracking HookedTransformer-era models on the bridge; lit/model.py's token-id positional indexing fixed.
Re-executed where executable – committed output conventions preserved (stripped notebooks stay stripped), metadata byte-identical; Activation_Patching's committed gpt2 numbers reproduce byte-for-byte on the bridge.
Deprecation mechanics

HookedRootModule + HookPoint kept – scrubbed of HookedTransformer references, un-deprecated, pinned by a keeper-contract test; Main_Demo's arbitrary-module pedagogy survives 4.0 verbatim.
PEP-562 lazy legacy names – import transformer_lens loads none of the deprecated stack; a de-shadowing module class makes class-vs-submodule resolution order-independent; TYPE_CHECKING bindings keep mypy exact.
Protocol typing – tools/training.py against a new TrainableTransformerLensModel (standalone, exactly the surface the loop touches), SVDInterpreter against the base protocol; both shaped around beartype's getattr_static protocol checking.
filterwarnings policy synced – newer class warnings added, obsolete HookedRootModule entry removed.

Type of change

• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• 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