fix: address M3 verifier findings on soak invariants and probe tests

tools/soak/scenarios.py

@@ -47,19 +47,31 @@ class ChsExtraction(BaseModel):
 
 @dataclass
 class SoakCorpus:
-    """一次 run 的语料句柄(装载一次,场景间共享)。"""
+    """一次 run 的语料句柄(装载一次,场景间共享)。
+
+    modalities 控制装载面: P7 只需 images——chains/replays 含整段 base64
+    负载,全量装载曾使 P7 起跑 RSS 冲到 ~690MB(verifier I1,2026-07-22)。
+    """
 
     harness_db: Path
     telemetry_db: Path
     frames_root: Path
     images_root: Path
+    modalities: tuple[str, ...] = ("chains", "replays", "images")
 
     def __post_init__(self) -> None:
-        self.chains = load_trace_chains(self.harness_db)
-        self.replays = load_replay_payloads(self.telemetry_db)
-        self.images = sorted(Path(self.images_root).glob("chs_*.jpg"))
-        if not (self.chains and self.replays and self.images):
-            raise ValueError("语料不完整: 请确认 data/soak/ 已按 findings §6 拉取")
+        self.chains = load_trace_chains(self.harness_db) if "chains" in self.modalities else []
+        self.replays = (
+            load_replay_payloads(self.telemetry_db) if "replays" in self.modalities else []
+        )
+        self.images = (
+            sorted(Path(self.images_root).glob("chs_*.jpg"))
+            if "images" in self.modalities
+            else []
+        )
+        missing = [m for m in self.modalities if not getattr(self, m)]
+        if missing:
+            raise ValueError(f"语料不完整({missing}): 请确认 data/soak/ 已按 findings §6 拉取")
 
 
 def weighted_mix(weights: dict[str, float], rng) -> str: