diff --git a/src/ghidra/GhidraProvider.behavior.test.ts b/src/ghidra/GhidraProvider.behavior.test.ts
index 8239426e5a..c2b958555f 100644
--- a/src/ghidra/GhidraProvider.behavior.test.ts
+++ b/src/ghidra/GhidraProvider.behavior.test.ts
@@ -3819,7 +3819,7 @@ const failureOutput = (stdout = "", stderr = "") => ({
   }),
 });
 const projectFailure = async (
-  kind: "cancelled" | "process" | "protocol" | "remote" | "timeout",
+  kind: "cancelled" | "process" | "protocol" | "remote" | "start" | "timeout",
   diagnostics: Readonly<Record<string, JsonValue>>,
 ) => {
   const ghidra = provider(installationHost(), () => ({
@@ -4161,6 +4161,19 @@ describe("Ghidra SIGKILL projection", () => {
     },
   );
 
+  it("keeps start-failure SIGKILL cause unknown despite a cgroup OOM delta", async () => {
+    const diagnostics = { ...killed, memory_cgroup: cgroup(1, 3_221_225_472) };
+    const projected = await projectFailure("start", diagnostics);
+    expect(projected.code).toBe("execution_failure");
+    expect(projected.message).toContain("SIGKILL");
+    expect(projected.message).toContain("does not establish its cause");
+    expect(projected.details?.diagnostics).toMatchObject(diagnostics);
+    expect(projected.remediation.action).toContain(
+      "host or container termination records",
+    );
+  });
+
+
   it.each([
     ["without cgroup evidence", {}],
     [
@@ -4175,9 +4188,18 @@ describe("Ghidra SIGKILL projection", () => {
     });
     expect(projected.code).toBe("execution_failure");
     expect(projected.message).toContain("SIGKILL");
-    expect(projected.message).toContain("out of memory");
+    expect(projected.message).toContain("does not establish its cause");
     expect(projected.message).not.toContain("rea doctor");
-    expect(projected.remediation.action).toContain("decompiler");
+    expect(projected.remediation.action).toContain(
+      "host or container termination records",
+    );
+    expect(projected.remediation.action).not.toContain(
+      "GHIDRA_HEADLESS_MAXMEM",
+    );
+    expect(projected.details?.diagnostics).toMatchObject({
+      ...killed,
+      ...extra,
+    });
   });
 
   it("keeps an observed OOME primary over the kill", async () => {
diff --git a/src/ghidra/GhidraProviderClient.ts b/src/ghidra/GhidraProviderClient.ts
index eb5ea3b1bf..f0cdbbf392 100644
--- a/src/ghidra/GhidraProviderClient.ts
+++ b/src/ghidra/GhidraProviderClient.ts
@@ -1025,7 +1025,7 @@ const projectPrimarySessionError = (
       ? "Ghidra's output for this operation is incomplete. No OutOfMemoryError was confirmed; a memory cause cannot be ruled out."
       : undefined;
   if (memoryFailure !== undefined) return memoryFailure;
-  const killed = projectExternalSigkill(operation, failure, context, {
+  const killed = projectSigkill(operation, failure, context, {
     restart: exitedProviderMessage(operation, failure, health),
     uncertainty,
   });
@@ -1049,18 +1049,6 @@ const projectPrimarySessionError = (
           }),
     });
   }
-  if (
-    (failure.kind === "start" ||
-      failure.kind === "process" ||
-      failure.kind === "protocol") &&
-    failure.diagnostics.exit_signal === "SIGKILL"
-  )
-    return new ProviderAdapterError("ghidra", operation, {
-      ...context,
-      userMessage: `Ghidra process diagnostics report termination by SIGKILL. The signal alone does not establish its cause or an out-of-memory failure.${uncertainty === undefined ? "" : ` ${uncertainty}`}`,
-      remediationAction:
-        "Review the retained process output and host or container termination records before choosing a resource-specific remedy. Export useful Evidence, close the binary and resolve any cleanup failure, then call open_binary for the same target to start a new Ghidra session. Session-only annotations are not retained.",
-    });
   const exited = exitedProviderMessage(operation, failure, health);
   if (exited !== undefined)
     return new ProviderAdapterError("ghidra", operation, {
@@ -1196,13 +1184,13 @@ const GHIDRA_PROCESS_TREE_MEMORY_ACTION =
   "A memory limit covers Ghidra's native decompiler processes as well as its Java heap, and their number follows the available CPUs, so a larger GHIDRA_HEADLESS_MAXMEM does not help. Raise the container or host memory limit, or allow fewer CPUs with a longer REA_GHIDRA_STARTUP_TIMEOUT_MS.";
 
 /**
- * A SIGKILL observed before the process supervisor began stopping Ghidra did
- * not come from REA. Kernel and container OOM killers send it without any
- * OutOfMemoryError text, so a rise in the oom_kill count of REA's cgroup is
- * the memory evidence when available. That count covers every process in the
- * cgroup and its child groups, not only REA's process tree.
+ * Own SIGKILL interpretation for startup and running-provider failures.
+ * The signal alone leaves its cause unknown. For process/protocol failures
+ * observed before REA requested a stop, a rising cgroup oom_kill count supplies
+ * separate memory evidence. That count covers the cgroup and its child groups,
+ * not only REA's process tree.
  */
-const projectExternalSigkill = (
+const projectSigkill = (
   operation: AnalysisOperation,
   failure: GhidraSessionError,
   context: {
@@ -1215,9 +1203,10 @@ const projectExternalSigkill = (
   },
 ): AnalysisError | undefined => {
   if (
-    (failure.kind !== "process" && failure.kind !== "protocol") ||
-    failure.diagnostics.exit_signal !== "SIGKILL" ||
-    failure.diagnostics.stop_requested !== false
+    (failure.kind !== "start" &&
+      failure.kind !== "process" &&
+      failure.kind !== "protocol") ||
+    failure.diagnostics.exit_signal !== "SIGKILL"
   )
     return undefined;
   const retry = guidance.restart ?? "Then retry.";
@@ -1225,6 +1214,8 @@ const projectExternalSigkill = (
     failure.diagnostics.memory_cgroup,
   );
   if (
+    (failure.kind === "process" || failure.kind === "protocol") &&
+    failure.diagnostics.stop_requested === false &&
     cgroup.success &&
     cgroup.data.oom_kill_after > cgroup.data.oom_kill_before
   ) {
@@ -1246,8 +1237,9 @@ const projectExternalSigkill = (
   }
   return new ProviderAdapterError("ghidra", operation, {
     ...context,
-    userMessage: `Ghidra was killed by SIGKILL before REA stopped it. REA could not confirm the sender; a kernel or container out of memory killer is a common one and leaves no Java OutOfMemoryError in the output.${guidance.uncertainty === undefined ? "" : ` ${guidance.uncertainty}`}`,
-    remediationAction: `Check the host or container for an out-of-memory kill, for example in the kernel log or as OOMKilled in docker inspect. ${GHIDRA_PROCESS_TREE_MEMORY_ACTION} ${retry}`,
+    userMessage: `Ghidra process diagnostics report termination by SIGKILL. The signal alone does not establish its cause or an out-of-memory failure.${guidance.uncertainty === undefined ? "" : ` ${guidance.uncertainty}`}`,
+    remediationAction:
+      "Review the retained process output and host or container termination records before choosing a resource-specific remedy. Export useful Evidence, close the binary and resolve any cleanup failure, then call open_binary for the same target to start a new Ghidra session. Session-only annotations are not retained.",
   });
 };