[Proposal] Migration recipes for the new bridge capabilities in migrating_to_v3.md

[jlarson4]

Proposal

Document the recently-added bridge capabilities in docs/source/content/migrating_to_v3.md, extending the existing ## Migrating specific HookedTransformer APIs section (heading at :211, table at :221-225): start_at_layer/input_to_embed, caching hooks (get_caching_hooks/add_caching_hooks), pos_slice, incl_bwd, HF-native past_key_values/logits_and_cache, mod_dict, reset_hooks parameters, and the model_protocol typing. Make sure you're based on dev-4.x, where these changes exist

Motivation

These primitives shipped on TransformerBridge with tests but zero docs. The migration guide's ## Migrating specific HookedTransformer APIs section explicitly invites new rows ("when you (or we) work out the equivalent, add a row below"), but currently holds only three rows (W_pos, W_E_pos, from_pretrained_no_processing).

Pitch

• Add one table row per capability to the table at migrating_to_v3.md:221-225, and for the four that need real code, a short prose recipe beneath the table (line 225 is currently end-of-file): start_at_layer resume-from-layer, caching hooks (+pos_slice/incl_bwd), past_key_values/logits_and_cache incremental decoding, and model_protocol typing for downstream libraries.
• Also add a config-construction row: HookedTransformerConfig(...) → TransformerBridgeConfig(...) — construct the bridge config directly with the same core fields; legacy config objects are deliberately not accepted by boot_native (it raises a clear TypeError — draft Attn only #20). This is the train-from-scratch on-ramp for the toy-model demos.
• Also add a calling-convention row: model.all_head_labels() (HT method) → bridge.all_head_labels (property — drop the parens). The bridge deliberately does not add a callable shim; calling the property raises TypeError: 'list' object is not callable, which this row preempts.
• Also add a tokenizer row: model.set_tokenizer(tok) → not ported; the tokenizer is fixed at boot — pass tokenizer= to boot_transformers(...) or re-boot. Warn that raw bridge.tokenizer = tok assignment skips the d_vocab/BOS wiring (see draft Fix bug in bwd hooks #17).
• Copy snippets from the shipped integration tests rather than inventing them.
• Write these four caveats verbatim next to the relevant recipe: start_at_layer excludes embedding-stage hooks and still runs blocks 0..k-1 on a discarded path; incl_bwd needs a scalar output (return_type="loss") and the transformers driver only; the KV-cache is HF's own cache object, not TransformerLensKeyValueCache; the caching-hooks recipe should steer users to get_caching_hooks/add_caching_hooks/run_with_cache — the bridge's cache_all/cache_some mirrors emit DeprecationWarning (bridge_core.py:705,727).
• Cross-link existing anchors where relevant: ### Will my numbers match HookedTransformer? (:84), ## Hook names (:109), ## APIs that are unchanged (:132).

Acceptance:

• Each capability has a table row; the four code-bearing ones have runnable HT-to-bridge recipes
• The four caveats appear next to the relevant recipe
• /build-docs succeeds; doctests pass
• make unit-test passes
• uv run mypy . passes

Mirror: transformer_lens/model_bridge/bridge_core.py:627 (get_caching_hooks; add_caching_hooks at :669); transformer_lens/model_bridge/transformer_bridge.py:1358 (start_at_layer param; implementation _setup_start_at_layer at :1656); transformer_bridge.py:1362 (past_key_values), :1618 (logits_and_cache); bridge_core.py:129 (mod_dict), :585 (reset_hooks); transformer_lens/model_protocol.py:24,:45. Tests to copy snippets from: tests/integration/model_bridge/test_bridge_{start_at_layer,caching_hooks,manual_kv_cache,kv_cache_parity}.py.

Checklist

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