phase 0-2: HF+PEFT pipeline, smoke, subspace alignment

Rip Axolotl/vLLM, switch to HF+PEFT functional pipeline.
Add LoRA/DoRA/PiSSA/DeLoRA train, delta-W diff, weight_steer hook,
sycophancy logratio eval, and SVD top-k + weak-readout alignment.
Smoke runs end-to-end on tiny-random qwen3 with BEARTYPE=1.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

scripts/adapter_sweep.py

@@ -1,5 +0,0 @@
-"""Phase 3 entrypoint: run replicate.py for each adapter in {lora, dora, pissa, delora}.
-
-Final output: a polars table with columns
-    (adapter, behavior, alpha, eval_logratio, subspace_alignment_ratio, train_time_s)
-"""

scripts/replicate.py

@@ -1,5 +0,0 @@
-"""Phase 1 entrypoint: data -> train pos -> train neg -> diff -> eval.
-
-Usage:
-    uv run python -m scripts.replicate --model Qwen/Qwen3-0.6B --behavior sycophancy --adapter lora
-"""

scripts/subspace_align.py

@@ -1 +0,0 @@
-"""Phase 2 entrypoint: project w onto SVD + AntiPaSTO subspaces, print alignment table."""