fix(m5): eval pipeline — anomaly_regions schema + nested JSON + pure-LLM baseline

Critical eval-pipeline bugs found while running the stage-2 model:

1. parse_anomaly_segments only recognized {"segments\:[[s,e]]}; the training
   data (data/instruct.py) emits {"anomaly_regions":[{"start","end"}]}.
   The model learned the correct format but scored 0 → VUS-PR=0. Now accepts
   both schemas + a balanced-brace JSON scanner (the flat regex could not
   match nested objects the model emits).
2. qa_judge._extract_json had the same nesting bug — reuse the scanner.
3. Generation truncated at 64-96 tokens (too short for stats JSON) → bumped
   to 192 in report + instruct_check.

Report enhancements:
- wire the pure-LLM baseline (series-as-text, no TS modality) so design §5.4
  'model vs pure LLM' is actually computed. Heavy predictors built/evaluated
  one at a time with GPU cleanup (single 12GB card).
- add QA-judge table across all 6 categories (design §6.5 问答均分).
- add --pure_llm / --max_per_cat flags.

Results (reports/eval-2026-06-30.md): model QA-judge mean 2.45 vs pure_llm
0.88 (2.8x); anomaly localization (VUS-PR) is the weak point — neither
model nor pure_llm localize well; TS-modality ablation change_rate 0.98.

scripts/run_m4_m5_eval.sh

@@ -0,0 +1,37 @@
+#!/usr/bin/env bash
+# M4 + M5 exit-validation eval (T4.2 / T4.3 / T5.6).
+#
+# Runs (1) the instruction-following check (T4.2) and (2) the full eval report
+# (T5.5/5.6) including the trained model + trivial baselines, on the stage-2
+# checkpoint. Requires a free GPU (so run AFTER training finishes).
+#
+# Usage:
+#   scripts/run_m4_m5_eval.sh [stage2_ckpt]
+set -euo pipefail
+cd "$(dirname "$0")/.."
+export PYTHONPATH=src
+
+CKPT="${1:-checkpoints/stage2/stage2_final.pt}"
+if [ ! -f "$CKPT" ]; then
+  echo "[run_m4_m5_eval] checkpoint $CKPT not found"; exit 1
+fi
+
+echo "=== T4.2 instruction-following check (6 cats × 10) ==="
+.venv/bin/python -m tsmm.eval.instruct_check \
+  --data data/eval_synth.jsonl \
+  --per_cat 10 \
+  --stage2_ckpt "$CKPT" \
+  --judge_backend stub \
+  --out reports/instruct_check.json
+
+echo
+echo "=== T5.5/T5.6 full eval report (model + trivial baselines) ==="
+.venv/bin/python -m tsmm.eval.report \
+  --synth data/eval_synth.jsonl \
+  --model \
+  --stage2_ckpt "$CKPT" \
+  --judge_backend stub \
+  --report_dir reports
+
+echo
+echo "[run_m4_m5_eval] done. See reports/."