refactor: split submit_sft_via_cml.sh into submit_sft.sh + submit_cml_eval.sh

- submit_sft.sh: pure SFT submission, prints JobID and exits (no embedded watcher)
- submit_cml_eval.sh: pure CML eval, reads workflow_id/version from config.yaml
- jupyter_runtime write_text: rewrite to use Contents API instead of terminal websocket
- SKILL.md: restrict trigger to explicit UI click only
- program.md: update §5.2 docs to reflect two-script workflow

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

backend/api/server.py

@@ -7537,6 +7537,30 @@ def _compute_kpis(
         if run_max is not None
         else None
     )
+    # Fallback: max runDic 可能命中 R{n} retry 轮（cml 还在 running，lark_template
+    # 没产出），让 KPI 全部空白。这时回退到 state 里最新一个 `cml=complete` 的
+    # 那一轮的 runDic — baseline / 上一轮成功 eval 的指标至少能显示出来。
+    if metrics is None:
+        per_round = _resolve_run_dic_per_round(state_entries)
+        for entry in reversed(state_entries):
+            if entry.get('step') != 'cml' or entry.get('status') != 'complete':
+                continue
+            rid = entry.get('run_id')
+            if not isinstance(rid, str):
+                continue
+            fallback_run = per_round.get(rid.strip())
+            if fallback_run is None or fallback_run == run_max:
+                continue
+            metrics = _read_workflow_metric_diff(
+                fallback_run,
+                agent_state=agent_state,
+                account_id=account_id,
+                session_id=session_id,
+            )
+            if metrics is not None:
+                run_max = fallback_run
+                run_id_value = str(fallback_run)
+                break
     _enrich_metrics_from_eval_output(metrics, target_set)
     target_metric = _format_metric_value(metrics, 'target_set_pass_rate')
     overall_metric = _format_metric_value(metrics, 'overall_car_pass_rate')