Real-GPU memory-non-growth validation for #671 hook-based extraction

kind: infra. Parent: #671 (the activation-extractor output_hidden_states memory-accumulation fix).

Goal

Confirm resident GPU memory stays flat across N forward passes on a real GPU — the authoritative check the #671 CPU proxy cannot do.

Why this is a separate task

#671 replaced subset-of-layers output_hidden_states=True reads with register_forward_hook + output_hidden_states=False across the live extraction line (scripts/issue667_extract.py, src/explore_persona_space/experiments/issue_650/shift_extract.py, src/explore_persona_space/analysis/probes.py, scripts/i488_phase1_predictors.py), routing them through the new shared src/explore_persona_space/analysis/extraction.py::extract_layer_activations.

The #671 acceptance gate was CPU-only: bit-for-bit numerical identity (tests/test_issue671_extraction_hooks.py) + an AST regression-lock + an advisory CPU ru_maxrss non-growth proxy. The real bug was GPU-segment retention under the expandable_segments:True allocator (#545 HF resident grew 22→30→38 GiB across rounds; #667 GCP networking wedge from memory pressure). That allocator behavior is a CUDA property absent on CPU — the CPU proxy can only confirm no Python-level reference is retained, a necessary-but-not-sufficient check. This task is the sufficient one.

Validation recipe (cheap, ~0.25 GPU-h)

1. Provision an eval-intent pod (1× H100) or use the auto GCP eval lane.
2. Load Qwen/Qwen2.5-7B-Instruct on GPU in eval mode.
3. Run N≈50 repeated extract_layer_activations(model, ids, [7, 14, 21]) calls on a fixed prompt (the live subset the #545/#651/#667 line reads), recording torch.cuda.memory_allocated() / torch.cuda.memory_reserved() after each.
4. Assert reserved memory does NOT grow monotonically across iterations (flat after a small warm-up), with PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True set (the allocator config under which #545 grew) AND once without it, so the read is attributable to the allocator regime.
5. As a positive control, run the SAME loop with the OLD model(ids, output_hidden_states=True) subset read and confirm IT grows (or at least retains more reserved memory) — demonstrating the fix removed a real growth, not just that the new path happens to be flat.

Done = reserved memory flat across iterations on the hook path, and the old-path positive control shows the growth the fix removed. Report the two memory curves.

Scope

Pure validation — no code changes expected (the fix already landed in #671). If the real-GPU read shows growth on the hook path, that is a #671 regression to root-cause, not a finding to file elsewhere.