artifacts: direction driver + directional ablation (Phase 0f)

Overview / Motivation

Phase 0f of the unified artifact factory (plan: /home/thomasjiralerspong/.claude/plans/help-me-to-devise-vectorized-tower.md). Build the behavior direction-vector driver + the missing directional-ablation path (the Arditi second half).

Goal

• src/explore_persona_space/artifacts/directions.py — the direction driver. Given a Behavior: generate contrastive exhibit/not-exhibit completions from the behavior's extraction_prompt_pairs (the 5 pairs, REQUIRED DISJOINT from the training instructions) — Claude-generated, judge-filtered (>50 / <50, drop-never-coerce via eval/graded_judge) — then teacher-force the kept completions through Qwen using analysis/extraction.extract_layer_activations (OOM-safe hook; NEVER output_hidden_states=True), take response-avg activations per layer, and compute diff-of-means r_B per layer (all 28 layers). Reuse scripts/issue779_extract_rb diff-of-means math. Enforce the regime: steering → pick ONE layer by steering effect; read_out → sweep all layers + the selection-symmetric-null harness (REFUSE a single-layer headline in read_out without per-draw same-selection OR a held-out-frozen axis; persist the per-draw × per-layer matrix). For 0f, accept a contrastive completion set as input (the datagen 0d path produces it later) so the module + tests land now; the full end-to-end run is exercised in Phase 2.
• src/explore_persona_space/artifacts/ablation.py — NEW directional ablation h ← h − (ĥ·r̂)r̂ via register_forward_hook per block (mirror scripts/issue623_extract_sycophancy_vector.py::steering_generate's hook structure but PROJECT-OUT instead of add); single-layer AND all-layer variants.

Scope / constraints

• Reuse extract_layer_activations, the issue779 diff-of-means math, eval/graded_judge. The teacher-forced-Claude-completions direction is a NAMED persona-vectors deviation (document in the plan §-assumptions per replication-fidelity.md / persona-vectors-recipe.md).
• Append package exports to artifacts/__init__.py append-only (sibling 0c/0e land concurrently).
• CPU/mock unit tests: ablation is a NO-OP on a zero direction (≡ base generation); diff-of-means on synthetic activations; the regime guard REFUSES a single-layer read_out headline without a symmetric null; per-draw × per-layer matrix persisted. A tiny GPU smoke for the teacher-force path is OK but keep it minimal. Lint + pytest.

Rules to read

.claude/rules/persona-vectors-recipe.md (7 steps, steering-vs-readout regime, the deviation), .claude/rules/selection-symmetric-nulls.md, .claude/rules/marker-leakage-measurement.md, .claude/rules/llm-judging.md.