Fix activation-extractor output_hidden_states memory accumulation (root cause of GCP DHCP-wedge + recurring #545 OOM): register_forward_hook on read layers + local-SSD .npz scratch

Problem

The activation extractor calls model(ids, output_hidden_states=True) on both base and trained models per (persona, question) while reading only layers 7/14/21 (scripts/issue667_extract.py:419-420, and the same pattern across the #545/#651 extraction line). This materializes all 29 residual-stream tensors × 2 models per forward; with the expandable_segments:True allocator retaining freed segments, resident memory climbs iteration-to-iteration (documented: #545 went 22→30→38 GiB across 5 rounds). Two consequences:

1. Recurring GPU OOM — #545 burned 5 implementer rounds on gpu_memory_utilization / batch tuning before an architectural pivot fixed the real cause.
2. Root cause of the GCP networking wedge (#667) — the climbing memory pressure drives kernel page-reclaim stalls that starve the systemd-networkd DHCP renewal → Could not set DHCPv4 address: Connection timed out → the hung-but-RUNNING zombie (deep-research run wf_857f7cfd; #669 is the recovery backstop, THIS task removes the trigger).

Fix (per agent-memory experiment-implementer/feedback_output_hidden_states_activation_accumulation_oom.md + in-repo precedent)

1. Extract via register_forward_hook on ONLY the needed layers (7/14/21), pass output_hidden_states=False, with per-iteration del captured + torch.cuda.empty_cache(). In-repo precedent: analysis/representation_shift.py:70-79.
2. Layer-index off-by-one (critical): out.hidden_states[k] = output of block k-1 (hs[0] = embed_tokens). A naive hook on model.model.layers[layer] captures hs[layer+1]. Map: layer 0 → embed_tokens; layer k≥1 → layers[k-1]. The hook path MUST reproduce the current output_hidden_states[layer+1] read bit-for-bit.
3. Keep a full-tuple fallback for non-standard models / CPU test stubs (if getattr(model.model, "layers", None) is None).

Secondary — GCE I/O decoupling (lower priority)

4. Write per-cell .npz to a local SSD scratch dir (not the network-backed GCE Persistent Disk) and batch-upload at cell/sweep end. GCE Persistent Disk I/O traverses the network, so heavy .npz writes saturate the same virtual NIC the DHCP renewal needs — the GCE-specific amplifier (deep-research wf_857f7cfd).

Tests (required)

• Bit-for-bit equivalence: a CPU test pinning hook-path output == the output_hidden_states[layer+1] path on a hook-capable stub (exposing model.model.layers / embed_tokens whose modules fire registered hooks).
• AST test that the primary read path uses register_forward_hook + output_hidden_states=False (no regression to output_hidden_states=True).
• Memory-non-growth assertion across N iterations if feasible.

Scope / acceptance

• Numerically identical activation reads (bit-for-bit test gates it); no resident growth across iterations.
• Generalize the fix to the shared extraction path where feasible (the same pattern bit #545/#651), not only issue667_extract.py.
• kind: infra — code-change path, 0 GPU (CPU stub tests), auto-completes on test-verdict PASS.

Provenance

Root-caused from the #667 GCP networking-wedge investigation + deep-research (wf_857f7cfd, 2026-06-25): the dominant trigger of the GCP hung-RUNNING zombie is this extractor memory bug, which is also the recurring #545 OOM. #669 is the recovery backstop; this task removes the trigger.