AGC Rust 分片失败复核改按名单精确匹配，覆盖 bin 目标用例名

- run-rust-shell-test-shards.mjs：extractFailingTestNames 改为按本片名单精确匹配用例名，不再用 `tests::` / `process_session::tests::` 前缀白名单——bin 目标的 `agent::…::tests::…` 原先永远进不了复核，偶发假红直接判红
- run-rust-shell-test-shards.mjs：复核也失败时同时保留首次失败现场，不再只留复核那一次
- run-rust-shell-test-shards.test.mjs：新增「命名空间用例名也会进复核」的回归用例
- 验证：node --test apps/ai-game-creator-shell/scripts/run-rust-shell-test-shards.test.mjs 3 项通过

apps/ai-game-creator-shell/scripts/run-rust-shell-test-shards.mjs

@@ -447,7 +447,12 @@ async function runWithConcurrency(shards, runner) {
   return results;
 }
 
-function extractFailingTestNames(result) {
+// 只认**本片名单里真实存在**的用例名：先按 `---- <name> stdout ----` / 失败清单行归一化，
+// 再拿名单做精确匹配。不要再用名字前缀白名单猜用例（历史上只放行 `tests::` 与
+// `process_session::tests::`，而 bin 目标的用例名形如 `agent::thread_manager::tests::…`，
+// 于是绝大多数用例永远进不了复核，偶发假红就直接判红）。
+function extractFailingTestNames(result, shardTestNames) {
+  const known = new Set(shardTestNames);
   const names = new Set();
   for (const line of result.failures) {
     const normalized = line
@@ -457,11 +462,8 @@ function extractFailingTestNames(result) {
       .replace(/\s*stdout\s*----$/, '')
       .replace(/\s*\(\d+\)\s*$/, '')
       .trim();
-    const match = normalized.match(
-      /^(process_session::tests::[A-Za-z0-9_:]+|tests::[A-Za-z0-9_:]+)$/,
-    );
-    if (match) {
-      names.add(match[1]);
+    if (known.has(normalized)) {
+      names.add(normalized);
     }
   }
   return [...names];
@@ -518,9 +520,7 @@ async function main() {
       }
       // 片内串行的时序型用例在高负载 CI 上会偶发假红。只对失败用例做一次
       // 有界复核：复核通过按 flaky 记录，复核失败才判红，避免把真实回归洗掉。
-      const failingTestNames = extractFailingTestNames(result).filter((name) =>
-        shardTestNames.includes(name),
-      );
+      const failingTestNames = extractFailingTestNames(result, shardTestNames);
       if (failingTestNames.length === 0) {
         return result;
       }
@@ -531,10 +531,13 @@ async function main() {
         failingTestNames,
       );
       if (!retry.ok) {
+        // 复核也失败就判红，但**两次的现场都留**：首次是整片的上下文，复核只跑失败用例，
+        // 丢掉首次现场会让"到底整片跑成什么样"这件事在日志里消失。
         return {
           ...result,
           failures: [
             `re-run of ${failingTestNames.length} failing test(s) also failed`,
+            ...result.failures,
             ...retry.failures,
           ],
         };

apps/ai-game-creator-shell/scripts/run-rust-shell-test-shards.test.mjs

@@ -68,3 +68,25 @@ test('successful shard keeps its compact summary', (t) => {
   assert.match(result.output, /shard 1\/1 ok: 1 test\(s\)/);
   assert.doesNotMatch(result.output, /test passing_case \.\.\. ok/);
 });
+
+// bin 目标的用例名带命名空间（`agent::thread_manager::tests::…`）。复核名单必须靠
+// 「本片名单精确匹配」认用例，不能靠名字前缀白名单，否则这类用例永远进不了重跑。
+test('namespaced failing test is re-run once before the shard is judged', (t) => {
+  const result = runFixture(
+    t,
+    `
+mod agent {
+    pub mod thread_manager {
+        #[test]
+        fn flaky_case() {
+            assert_eq!(1, 2, "namespaced panic evidence");
+        }
+    }
+}
+`,
+  );
+  assert.equal(result.status, 1, result.output);
+  assert.match(result.output, /re-run of 1 failing test\(s\) also failed/);
+  assert.match(result.output, /agent::thread_manager::flaky_case/);
+  assert.match(result.output, /namespaced panic evidence/);
+});