extraction.py docstring: last-layer hook capture is pre-final-norm (not identical to hidden_states[-1])

Overview / Motivation

Filed from a "Needs human eyeball" prose follow-up in #779 round-6 implementation (experiment-implementer): a documentation inaccuracy in a shared src/ helper that could mislead future capture work.

Goal

Fix the extract_layer_activations hook-path docstring in src/explore_persona_space/analysis/extraction.py: the LAST-layer hook capture returns the RAW final-block residual (pre-final-norm), while output_hidden_states[-1] is POST-final-norm — the "#671 byte-identical to the full-tuple read" claim is stub-true (tiny random models) but real-model-FALSE for the last layer specifically.

Detail

• Non-last layers ARE identical between the hook path and hidden_states[idx].
• Only the last decoder block's hook output differs from hidden_states[-1] by the final RMSNorm.
• Pre-existing (NOT introduced by the #779 round-6 fix); pre/post-patch behavior identical (proved by the round-6 equivalence smoke). #779's own use is internally consistent (all its extractions ride the same hook convention).
• Fix scope: docstring + (optionally) a one-line comment at the hook registration; NO behavior change. Consider a short unit-test comment in tests/test_issue671_extraction_hooks.py clarifying the convention.

Constraints

• Documentation-only; zero behavior change; ruff clean.