fix(wizard): read keypresses from the raw fd so arrow keys work

sys.stdin.read(1) on the buffered TextIOWrapper pulls the whole escape
sequence into its Python-level buffer, so the follow-up select() sees an
empty fd and every arrow/Home/End press degrades to bare Esc; the unread
'[B' tail then leaks into the next keypress. Read raw bytes via os.read()
and select() on the file descriptor instead, keeping the same key tokens
and classifications. Add pty-driven regression tests over a real cbreak
terminal covering CSI sequences, single-byte keys, bare Esc, and tail-leak.

src/weekly_activity/wizard.py

@@ -343,23 +343,26 @@ def _read_keypress() -> str:
     """Read one keystroke from stdin (cbreak mode); classify it."""
     import select
 
-    char = sys.stdin.read(1)
-    if char == "\x1b":
+    fd = sys.stdin.fileno()
+    # Read raw bytes from the fd: a buffered text read would pull the whole
+    # escape sequence into its own buffer, leaving nothing for select to see.
+    first = os.read(fd, 1)
+    if first == b"\x1b":
         # A bare Esc is not followed by more bytes; CSI sequences are. Peek briefly.
-        readable, _, _ = select.select([sys.stdin], [], [], 0.05)
+        readable, _, _ = select.select([fd], [], [], 0.05)
         if not readable:
             return _KEY_ESCAPE
-        tail = sys.stdin.read(2)
-        return _CSI_KEYS.get(tail, _KEY_OTHER)
-    if char in ("\r", "\n"):
+        tail = os.read(fd, 2)
+        return _CSI_KEYS.get(tail.decode("ascii", "replace"), _KEY_OTHER)
+    if first in (b"\r", b"\n"):
         return _KEY_ENTER
-    if char == "\x01":  # Ctrl-A
+    if first == b"\x01":  # Ctrl-A
         return _KEY_HOME
-    if char == "\x05":  # Ctrl-E
+    if first == b"\x05":  # Ctrl-E
         return _KEY_END
-    if char == "k":
+    if first == b"k":
         return _KEY_UP
-    if char == "j":
+    if first == b"j":
         return _KEY_DOWN
     return _KEY_OTHER