some changes and log fix

python_app/orchestration/process_supervisor.py

@@ -105,7 +105,10 @@ def _read_log_tail(path: Path, max_bytes: int = 16384) -> str:
             data = handle.read()
     except OSError:
         return ""
-    return data.decode("utf-8", errors="replace").strip()
+    # Drop NULs: logs written by an older supervisor can carry a sparse hole from
+    # the pre-O_APPEND truncate bug, and a tail landing in it would otherwise turn
+    # an exit report (or a rolled `.prev`) into megabytes of NUL padding.
+    return data.replace(b"\0", b"").decode("utf-8", errors="replace").strip()
 
 
 class ProcessSupervisor:
@@ -232,8 +235,17 @@ class ProcessSupervisor:
         self._roll_log_to_prev(stdout_path)
         self._roll_log_to_prev(stderr_path)
 
-        stdout_file = open(stdout_path, "wb")
-        stderr_file = open(stderr_path, "wb")
+        # O_APPEND ("ab"), not "wb": the child inherits these fds and keeps its own
+        # file offset. Without O_APPEND, the in-place truncate in
+        # `_roll_log_if_oversized` leaves that offset far past the new end of file,
+        # so the next write lands there and the kernel fills everything before it
+        # with a hole of NUL bytes — the log becomes unreadable and the size cap
+        # stops working entirely. O_APPEND makes the kernel seek to EOF atomically
+        # on every write, so a truncate genuinely restarts the file at offset 0.
+        # `_roll_log_to_prev` above already renamed any previous log away, so not
+        # truncating on open costs nothing.
+        stdout_file = open(stdout_path, "ab")
+        stderr_file = open(stderr_path, "ab")
         try:
             handle = subprocess.Popen(
                 command,
@@ -499,6 +511,11 @@ class ProcessSupervisor:
         The child holds an open fd to this inode, so a rename would not redirect
         its writes. Instead keep one rolled generation via copy-to-`.prev` and
         truncate the live inode in place, freeing the allocated disk blocks.
+
+        This relies on the child's fd being opened with O_APPEND (see `_spawn`):
+        only then does the child resume writing at offset 0 after the truncate.
+        With a plain write fd it would keep writing at its stale offset, punching
+        a multi-hundred-megabyte NUL hole and defeating the cap.
         """
         try:
             if path.stat().st_size <= _LOG_MAX_BYTES: