fix(sse): re-run strict system hoist after format translation (#10803)

Re-runs hoistLeadingSystemMessage on the final outbound array at translateRequest's single return, instead of only pre-translation. claudeToOpenAI (and the Responses source path, which never ran the pre-translation hoist at all since `messages` doesn't exist yet there) re-introduces/normalizes a leading system message after the hoist already ran, so a strict provider (e.g. vLLM/Qwen3, xiaomi-mimo) could still receive a non-compliant array and 400 with "System message must be at the beginning."

Validated live against a vLLM/Qwen3 endpoint (documented in the PR) plus in an isolated worktree boarded onto origin/release/v3.8.50 (0 conflicts, 2 files):
- 39/39 focused tests pass (probe-7293-strict-system-hoist including the new Claude-source regression case, memory-system-first-6135, claude-system-role-cache-boundary, memory-cache-safe-injection).
- check-file-size, check-changelog-integrity: OK.
- typecheck:core: clean.
- check-complexity / check-cognitive-complexity: OK, both under baseline.

Co-authored-by: Kizuno18 <Kizuno18@users.noreply.github.com>

open-sse/translator/index.ts

@@ -756,6 +756,19 @@ export function translateRequest(
     delete result[RESPONSES_STORE_MARKER];
   }
 
+  // #7293 follow-up: the pre-translation hoist above normalizes the *source*
+  // message array, which a target translator can then undo. `claudeToOpenAI`
+  // pushes `body.system` as a fresh leading system message before appending the
+  // converted messages, so an already-hoisted system lands at index 1 again;
+  // a Responses-source request has no `messages` at all until translation, so
+  // the earlier call is a no-op for it. Re-run on the final outbound array —
+  // it is the only shape the upstream actually sees. Idempotent: same array
+  // reference for non-strict providers and already-compliant requests, so
+  // prompt-cache prefixes stay stable.
+  if (targetFormat === FORMATS.OPENAI && result.messages && Array.isArray(result.messages)) {
+    result.messages = hoistLeadingSystemMessage(result.messages, provider);
+  }
+
   return result;
 }
 

tests/unit/probe-7293-strict-system-hoist.test.ts

@@ -142,3 +142,39 @@ test("#7293: already-compliant strict-provider request is a no-op (prompt-cache
 
   assert.deepEqual(result.messages, messages);
 });
+
+test("#7293: Claude-source request keeps a single leading system message after claudeToOpenAI re-adds body.system", () => {
+  // Claude Code's real shape: a top-level `system` field AND a system-role
+  // message inside `messages`. claudeToOpenAI pushes body.system as the leading
+  // system message and then appends the converted messages, so hoisting before
+  // translation is not enough — the offender reappears at index 1.
+  const body = {
+    model: "mimo-v2.5",
+    system: [{ type: "text", text: "You are a coding assistant." }],
+    messages: [
+      { role: "user", content: "hi" },
+      { role: "system", content: "deferred tools list" },
+      { role: "user", content: "go" },
+    ],
+  };
+
+  const result = translateRequest(
+    FORMATS.CLAUDE,
+    FORMATS.OPENAI,
+    "mimo-v2.5",
+    body,
+    false,
+    null,
+    "xiaomi-mimo"
+  );
+
+  const outMessages = result.messages as Array<{ role: string; content: string }>;
+  const systemIndices = outMessages
+    .map((m, i) => (m.role === "system" ? i : -1))
+    .filter((i) => i >= 0);
+
+  assert.deepEqual(systemIndices, [0]);
+  // Merge, never drop: both the top-level system and the offender survive.
+  assert.match(outMessages[0].content, /You are a coding assistant\./);
+  assert.match(outMessages[0].content, /deferred tools list/);
+});