stage-3c: working QLoRA training + eval — pytorch base, Qwen3.5 slug, SFTConfig

Training and eval now run clean on the unsloth 2026.5.2 / transformers v5 /
torch 2.10 stack. Fixes: pytorch/pytorch base image (sidesteps the nvidia/cuda
apt-signature failure and the torch download), correct base-model slug
unsloth/Qwen3.5-4B, TRL SFTConfig API. Adds scripts/eval_adapter.py — runs
dataset rows through base+adapter with structured (transformers-v5) message
content and Qwen3.5 thinking-mode stripping.

First v1 adapter: loss 2.10 -> 0.32 over 3 epochs. Eval surfaced an ill-posed
ioc_extraction dataset (output URL not present in input) — to be fixed in the
ExampleBuilder before the next training run.

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

README.md

@@ -121,7 +121,7 @@ To fine-tune Qwen3.5-4B with QLoRA in an NVIDIA Docker container:
 # 1. build datasets (one-off; re-run after ingestion changes)
 .venv/bin/psyc train-build-all
 
-# 2. build the training image (CUDA 12.4 + unsloth + Qwen3.5)
+# 2. build the training image (pytorch 2.6/CUDA 12.4 base + unsloth + Qwen3.5)
 docker build -t psyc-trainer -f Dockerfile.train .
 
 # 3. fine-tune (mount host data/ so adapters land there)
@@ -135,14 +135,25 @@ docker run --gpus all --rm \
     --output /data/adapters/psyc-v1
 ```
 
-Defaults target a 24 GB consumer GPU (3090/4090): Qwen3.5-4B-Instruct at 4-bit,
+Defaults target a 24 GB consumer GPU (3090/4090): `unsloth/Qwen3.5-4B` at 4-bit,
 LoRA `r=16`/`alpha=16`, bf16, 3 epochs, effective batch size 8. For A100-40/80
-bump `--base-model unsloth/Qwen3.5-9B-Instruct-bnb-4bit` and raise
-`--batch-size` + `--max-seq-length`.
+bump `--base-model unsloth/Qwen3.5-9B` and raise `--batch-size` +
+`--max-seq-length`.
 
 Output: `data/adapters/psyc-v1/final/` (adapter weights) + `training_meta.json`
 (base model, hyperparameters, dataset list).
 
+Evaluate the adapter against held-out dataset rows:
+
+```bash
+docker run --gpus all --rm \
+    --entrypoint python \
+    -v $(pwd)/data:/data -v $(pwd)/scripts:/scripts \
+    psyc-trainer /scripts/eval_adapter.py \
+    --adapter /data/adapters/psyc-v1/final \
+    --dataset /data/datasets/ioc_extraction-v1.jsonl --n 5
+```
+
 ## Status
 
 Day 2 of a 48h build. Shipped: Scoutline (URLhaus) → Classifyline → Mapline