refactor: drop the transformerlens path from graph /steer

[hijohnnylin]

Summary

Follow-up to 789942e, which fixed /steer on the interp_engine model engine. That fix kept a
second code path for transformerlens, and no graph pod runs that engine — so it was a second
untested path carrying a second spelling of every generation keyword.

• /steer now generates on interp_engine alone. generate_default and generate_steered read
  straight through with no engine dispatch, and _tl_kwargs is gone.
• The other engines are refused by name. _require_interp_engine separates two misconfigurations
  that used to share one message: a model reporting a backend this endpoint cannot use, and a
  model with no backend at all, which is what the lm-saes-crm attribution engine loads. A pod
  on another engine still generates graphs and fails only here.
• Nothing about the wire contract changes. temperature=0 still means greedy and freq_penalty
  still subtracts from a logit rather than scaling it, since that is what the steer modal sends, so
  _TransformerLensSampling and the translation into transformers' terms stay as they are.

--model-engine transformerlens remains available for graph generation, and the CRM parity harness
keeps its TransformerLens oracle — deleting that would remove the evidence behind defaulting to
interp_engine. README.md already documents /steer as the one narrower endpoint.

Test plan

• apps/graph: 11 unit tests in tests/test_steer_generation.py, covering the keyword
  translation, the frequency-penalty ordering, the token collector and each refused engine.
  They run on a randomly initialized two-layer GPT-2 built from a local config, so no
  download, no GPU, no tokenizer fetch.
• apps/graph: full suite, 66 passed / 10 skipped (the skips are the opt-in CRM parity tests).
• pyright at zero for the app; make python-lint clean.
• Real gemma-2-2b with the gemma-scope transcoders on a GPU, across all four combinations of
  freeze_attention and sampling: correct token counts, one row of logits per generated token,
  and steering visibly moves the continuation. Freezing changes the output at temperature 0.8,
  confirming the frozen attention patterns reach the backend.

Notes for review

The six-line change to the test file is renames only (_hf_kwargs -> _generation_kwargs), so
nothing it asserts moved — a reasonable signal that the deleted path was not carrying the substance.

Made with Cursor