fix(tests): un-quarantine encoder acceptance suites, sync quarantine inventory

[sohv]

Description

Two changes both making tests/QUARANTINES.md true again.

1. Un-quarantine the encoder acceptance suites

test_hooked_encoder.py and test_hooked_encoder_decoder.py were skipped in 3efbd6e ("Cleanup (#1129)", 2025-11-15) as "CI test pollution". Neither shows any order-dependence.

T5 — 22 passed, 1 skipped, 0 failed standalone and inside the full tier. No defect; the skip is simply removed.

BERT — 13 of 16 passed; the one failure was test_bert_block:

AssertionError: The values for attribute 'shape' do not match:
torch.Size([1, 8, 768]) != torch.Size([8, 768])

BertLayer.forward returned tuple[torch.Tensor] in transformers 4.56 and returns a bare torch.Tensor in 5.x, so the test's [0] went from unpacking the tuple to indexing the batch dimension. One line.

Flagging for review: the module was quarantined 2025-11-15, while the pin was still transformers>=4.56 — so this test was passing when it was skipped. It broke at the v5 floor bump in 8e8d9d4 (#1315) on 2026-05-19 and went unseen for ~2.5 months because the module was dark. The three sibling [0] unpacks in the same file (:58, :88, :108) are still correct — BertSelfAttention and BertEncoder kept their contracts.

Restores 40 tests (36 runnable, 4 environment-gated).

2. Sync the inventory with the files #1603 removed

Per the discussion on #1603 and #1605. That PR deleted 35 test files without touching tests/QUARANTINES.md, leaving 10 rows pointing at files that no longer exist:

Removed rows	Deleted file
×2	acceptance/test_hooked_transformer.py
×2	acceptance/test_multi_gpu.py
×1	integration/test_tensor_extraction_consistency.py
×1	integration/test_tokenization_methods.py
×3	unit/components/test_attention.py
×1	unit/factories/test_mlp_factory.py

With test_hooked_transformer.py gone, no whole-file quarantines remain, so the "⚠️ Technical debt — whole-file" section is dropped entirely.

AGENTS.md and docs/source/content/contributing.md both stated that the HookedTransformer acceptance suite was quarantined and named the two encoder files alongside it; both are corrected to reflect that the suite no longer exists and the encoder suites now run. Also fixes three drifted line numbers and notes that the multi-GPU validation record refers to a since-removed file.

Verification

• pytest tests/acceptance -m "not slow" → 230 passed, 31 skipped, 11 deselected, 0 failed
• Every path referenced in QUARANTINES.md now resolves to a file that exists
• Every module-level skip marker under tests/ is accounted for in the inventory
• pycln / isort / black no-ops; mypy clean

Notes for the maintainer

The two restored suites are cheap- BERT ~2.5 GB peak, T5 ~1.8 GB, ~35 s each. Having measured on the serial acceptance tier before #1603 landed, adding them changed peak RSS by less than run-to-run noise (7.53 GB without vs 6.26 GB with), so they should not affect the coverage-test job's memory headroom. Those figures are macOS-arm64 CPU and will differ on the Linux runners- compatibility-checks is the job that actually exercises the restored tests.

Not included here but raised separately in #1605: redwood_attn_2l is still listed in transformer_lens/supported_models.py:24 but cannot load on the current dependency floor — its tokenizer repo (ArthurConmy/redwood_tokenizer) ships a tokenizer.json with 31,856 of 51,743 merges referencing tokens absent from its own vocab, which tokenizers 0.22.1 rejects. The only test that exercised it was deleted by #1603, so this is now purely a model-support question.

Type of change

• Bug fix (non-breaking change which fixes an issue)
• 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 (restores 40 previously-dark tests; test_bert_block is red-before / green-after)
• New and existing unit tests pass locally with my changes