Previously, _drain() only captured the first MAX_STDOUT_BYTES (50KB) of
stdout, silently dropping all tail output. Scripts that print() their
final results at the end would have those results lost.
Now uses a two-buffer approach: 40% head + 60% tail (rolling window).
This matches the pattern already used in terminal_tool.py (line 1042-1051)
but gives the tail more space since execute_code scripts typically
print() their final results at the end.
Inspired by openclaw's softTrim context-pruning (headChars/tailChars).
