N1 — rename backend to engine, codebase-wide

[can-goodfire]

The nnsight-engine plan (cockpit notes/causalab-nnsight-engine-plan.md, §0 decision 6 / §6-N1) fixes the term engine for what protocol/backend.py already described as "one execution engine". This PR is the codebase-wide rename, landed first in the N-stack because it is only cheap pre-merge: the artifact stamp key rides along, and no artifacts are canon yet.

What renames

• protocol/backend.py → protocol/engine.py; Backend → Engine, choose_backend → choose_engine. ExecutionRequest/RunResult keep their names.
• neural/pytorch_hooks/ → neural/engines/pytorch_hooks/; backend.py → engine.py; PytorchHooksBackend → PytorchHooksEngine. A shim at the old package path re-exports for one deprecation beat.
• Stamp key identity_base["backend"] → "engine" (ARTIFACT_IDENTITY_KEYS, both stamping sites: workflow/runner.py, engines/pytorch_hooks/engine.py).
• tests/neural/pytorch_hooks/ → tests/neural/engines/pytorch_hooks/; test_backend_routing.py → test_engine_routing.py.
• Living docs (intervention_protocol.md §8 + capability table, workflow_protocol.md, CODEBASE.md, TESTS.md, READMEs), CLI refusal messages.

What deliberately does not

• The corpus fixture keeps the old stamp key. tests/protocol/_env.py generates 09's "previously fitted" artifact at test time; pre-rename artifacts factually carry "backend", and keeping it keeps 09's content_digest — hence the pinned canonical form — byte-stable, and keeps the loader's tolerance of pre-rename bundles under test. This is the one place the plan's "stamp is metadata, not canonical form" reasoning meets a fixture whose bytes embed the stamp.
• Historical docs (REBASE_CAUSALAB_ON_NNTERP.md, test_migration.md, NNsight_overview.md), pinned goldens JSON, and third-party senses (matplotlib Agg, LAPACK, uv build-backend, torch.backends).

Gates

• pytest -m "not golden": 1963 passed.
• update_corpus_digests.py: zero diff.
• pre-commit run --all-files: clean.
• grep gate: no [Bb]ackend outside the exemptions above.

Note for #54 (can/backend-forward-interning): it targets this base and touches the renamed executor; it will need a rebase over this (its CI is currently red anyway).

🤖 Generated with Claude Code

[github-actions]

Test Results

1 962 tests  ±0   1 954 ✅ ±0   58s ⏱️ -10s
    1 suites ±0       8 💤 ±0 
    1 files   ±0       0 ❌ ±0 

Results for commit 514a70b. ± Comparison against base commit 58be68e.

This pull request removes 365 and adds 365 tests. Note that renamed tests count towards both.

tests.neural.pytorch_hooks.test_attention_probs ‑ test_a_deltanet_layer_refuses_on_the_architecture
tests.neural.pytorch_hooks.test_attention_probs ‑ test_a_family_without_eager_math_refuses_by_name
tests.neural.pytorch_hooks.test_attention_probs ‑ test_a_generated_frame_read_refuses
tests.neural.pytorch_hooks.test_attention_probs ‑ test_a_non_swap_pattern_write_is_refused
tests.neural.pytorch_hooks.test_attention_probs ‑ test_an_edit_to_one_layer_leaves_the_other_layers_pattern_intact
tests.neural.pytorch_hooks.test_attention_probs ‑ test_an_identity_edit_is_bit_identical
tests.neural.pytorch_hooks.test_attention_probs ‑ test_an_identity_edit_is_bit_identical_on_llama
tests.neural.pytorch_hooks.test_attention_probs ‑ test_an_interchange_across_different_lengths_refuses
tests.neural.pytorch_hooks.test_attention_probs ‑ test_no_edits_installs_nothing
tests.neural.pytorch_hooks.test_attention_probs ‑ test_the_backend_now_declares_the_capability
…

tests.neural.engines.pytorch_hooks.test_attention_probs ‑ test_a_deltanet_layer_refuses_on_the_architecture
tests.neural.engines.pytorch_hooks.test_attention_probs ‑ test_a_family_without_eager_math_refuses_by_name
tests.neural.engines.pytorch_hooks.test_attention_probs ‑ test_a_generated_frame_read_refuses
tests.neural.engines.pytorch_hooks.test_attention_probs ‑ test_a_non_swap_pattern_write_is_refused
tests.neural.engines.pytorch_hooks.test_attention_probs ‑ test_an_edit_to_one_layer_leaves_the_other_layers_pattern_intact
tests.neural.engines.pytorch_hooks.test_attention_probs ‑ test_an_identity_edit_is_bit_identical
tests.neural.engines.pytorch_hooks.test_attention_probs ‑ test_an_identity_edit_is_bit_identical_on_llama
tests.neural.engines.pytorch_hooks.test_attention_probs ‑ test_an_interchange_across_different_lengths_refuses
tests.neural.engines.pytorch_hooks.test_attention_probs ‑ test_no_edits_installs_nothing
tests.neural.engines.pytorch_hooks.test_attention_probs ‑ test_the_engine_now_declares_the_capability
…