seccomp: require child memory peek in the availability probe

With Yama ptrace_scope=2 and no CAP_SYS_PTRACE, process_vm_readv and
/proc/<pid>/mem both fail, so sc_decide cannot read path strings and every
workspace write gets EACCES. The old probe only checked the USER_NOTIF
round trip, so CST_SANDBOX=seccomp and the forced-seccomp test pass looked
available and then failed on the first write.

sc_can_peek() forks a child, places a marker on its stack, and retries the
same sc_peek path. sc_probe() now needs that too, so those hosts fall back
to none (or keep landlock) instead of a silent deny-all sandbox.

README.md

@@ -279,6 +279,7 @@ TUI 中 `/tools off NAME`、`/tools on NAME` 可临时开关单个工具。
     （界面显示 `auto · seccomp`）。子进程的写类系统调用（带写标志的 open、创建 / 删除 / 改名 / 链接 / chmod 等）先交给
     agent 里的监督线程，按真实路径判断是否在工作区、临时目录或 `/dev/null` 等之内，不在就返回 `EACCES`；IPv4/IPv6 socket 一律拒绝。
     读、执行和已打开文件的写入不检查。它防的是误操作而不是恶意程序：检查之后内核会再读一次路径，多线程程序可以趁机换掉。
+    探测还会确认能读直接子进程的内存；`ptrace_scope=2` 且无 `CAP_SYS_PTRACE` 时 seccomp 不会启用。
     比 shell 活得久的后台进程：桌面发行版开着 Yama（`ptrace_scope` ≥ 1）时它们的受检写入会被拒；命令结束后一律失败。
   - macOS：`sandbox_init` profile，同样只放行工作区和临时目录写入，禁网。
   - Windows 10 1903+：Low 完整性级别的受限令牌 + Job Object（界面显示 `auto · lowil`）。第一次在工作区运行时

docs/compat.md

@@ -66,16 +66,19 @@ every create / remove / rename / link / chmod / chown / utimes / xattr call
 stop until the supervisor has read the path from the child's memory,
 resolved it (cwd, dirfd, `..`, symlinks, `/proc/self`, `/dev/stdout`) and
 answered: continue when it lands in the allowed places, `EACCES` otherwise.
-Reads, exec and writes to already-open files are not checked. Background
+Reads, exec and writes to already-open files are not checked. The probe also
-jobs that outlive their shell are no longer the agent's descendants: with
+checks that the parent can peek a direct child's memory (`process_vm_readv`
-Yama `ptrace_scope` ≥ 1 (common on desktop distros, absent on Android) the
+or `/proc/<pid>/mem`); with Yama `ptrace_scope=2` and no `CAP_SYS_PTRACE`
-supervisor cannot read their memory and their checked writes fail with
+that fails, so seccomp is not selected (forced `CST_SANDBOX=seccomp` then
-`EACCES`; once the run has finished they get `ENOSYS`. This keeps
+falls back to unavailable). Background jobs that outlive their shell are no
-mistakes out of the rest of the disk but is not a security boundary: the
+longer the agent's descendants: with Yama `ptrace_scope` ≥ 1 (common on
-kernel reads the path again after the check, so a hostile multi-threaded
+desktop distros, absent on Android) the supervisor cannot read their memory
-program can swap it in between. The availability check runs in a
+and their checked writes fail with `EACCES`; once the run has finished they
-throwaway child, so a platform seccomp policy that kills the caller cannot
+get `ENOSYS`. This keeps mistakes out of the rest of the disk but is not a
-take the agent down.
+security boundary: the kernel reads the path again after the check, so a
+hostile multi-threaded program can swap it in between. The availability
+check runs in a throwaway child, so a platform seccomp policy that kills the
+caller cannot take the agent down.
 
 ## Terminal notes
 

src/sandbox_exec.c

@@ -391,6 +391,47 @@ static ssize_t sc_peek(int pid, uint64_t addr, void *out, size_t n) {
   return k;
 }
 
+/* Path decisions need sc_peek on the sandboxed child. Yama
+ * kernel.yama.ptrace_scope=2 (and no CAP_SYS_PTRACE) denies that, so a
+ * listener-only probe would report seccomp available while every write
+ * fails with EACCES. */
+static int sc_can_peek(void) {
+  static const char want[] = "cst-peek-ok";
+  int sp[2];
+  if (pipe2(sp, O_CLOEXEC) != 0)
+    return 0;
+  pid_t pid = fork();
+  if (pid == 0) {
+    char marker[] = "cst-peek-ok";
+    uint64_t addr = (uint64_t)(uintptr_t)marker;
+    close(sp[0]);
+    if (write(sp[1], &addr, sizeof(addr)) != (ssize_t)sizeof(addr))
+      _exit(1);
+    close(sp[1]);
+    pause();
+    _exit(0);
+  }
+  close(sp[1]);
+  if (pid < 0) {
+    close(sp[0]);
+    return 0;
+  }
+  uint64_t addr = 0;
+  int ok = 0;
+  if (read(sp[0], &addr, sizeof(addr)) == (ssize_t)sizeof(addr) && addr) {
+    char got[sizeof(want)];
+    if (sc_peek((int)pid, addr, got, sizeof(got)) == (ssize_t)sizeof(got) &&
+        memcmp(got, want, sizeof(got)) == 0)
+      ok = 1;
+  }
+  close(sp[0]);
+  kill(pid, SIGKILL);
+  int st = 0;
+  while (waitpid(pid, &st, 0) < 0 && errno == EINTR)
+    ;
+  return ok;
+}
+
 static int sc_read_mem(int pid, uint64_t addr, void *out, size_t n) {
   size_t got = 0;
   while (got < n) {
@@ -862,7 +903,7 @@ static int sc_probe(void) {
   int st = 0;
   while (waitpid(pid, &st, 0) < 0 && errno == EINTR)
     ;
-  return ok && WIFEXITED(st) && WEXITSTATUS(st) == 0;
+  return ok && WIFEXITED(st) && WEXITSTATUS(st) == 0 && sc_can_peek();
 }
 
 #endif /* CST_HAVE_SECCOMP */

tests/test_sandbox.c

@@ -367,7 +367,19 @@ int main(void) {
       printf("test_sandbox: forced seccomp\n");
       test_confined(CST_SANDBOX_SECCOMP);
     } else {
-      printf("test_sandbox: seccomp not available here\n");
+      FILE *yf = fopen("/proc/sys/kernel/yama/ptrace_scope", "r");
+      int yama = -1;
+      if (yf) {
+        if (fscanf(yf, "%d", &yama) != 1)
+          yama = -1;
+        fclose(yf);
+      }
+      if (yama >= 2)
+        printf("test_sandbox: seccomp not available here "
+               "(yama.ptrace_scope=%d; cannot peek child memory)\n",
+               yama);
+      else
+        printf("test_sandbox: seccomp not available here\n");
     }
     if (was)
       setenv("CST_SANDBOX", saved, 1);