fix: rabbit review 001 - persist Streamlit failures, guard Jupyter lifecycle

- st_demo: persist failure in call history, clear active_handle,
  and disable Run while call active to avoid overwriting handle
- JupyterApp.shutdown: gather all shutdowns with return_exceptions=True
  and re-raise as BaseExceptionGroup so one failure doesn't leak kernels
- messages: fallback known-but-malformed messages to UnknownJupyterMessage
  via ValidationError, make Unknown channel permissive (str) for nuance
- transport: record reader failures as terminal state (_reader_failure)
  and fail fast from execute/messages_for instead of hanging
- shell: don't reset execution_count to 0 when IPython returns None
- README: use ipython_shell imports

Skipped non-urgent/deferred items: line(), get_shell_2, test renames,
import laziness, call-ID bounding, settle window config (worth doing later).

src/ipython_shell/jupyter/messages.py

@@ -1,6 +1,6 @@
 from typing import Annotated, ClassVar, Final, Literal
 
-from pydantic import BaseModel, ConfigDict, Field, TypeAdapter
+from pydantic import BaseModel, ConfigDict, Field, TypeAdapter, ValidationError
 
 
 type JupyterChannel = Literal["iopub", "shell", "stdin"]
@@ -176,6 +176,9 @@ class InputRequestMessage(_TypedMessage):
 class UnknownJupyterMessage(JupyterMessage):
     """An extension or future message preserved without lossy parsing."""
 
+    # Allow any channel value so known-but-malformed fallbacks never fail
+    # on a strict Literal. The parser preserves the transport channel verbatim.
+    channel: str = "iopub"  # type: ignore[assignment]
     raw: dict[str, object] = Field(default_factory=dict)
 
 
@@ -252,7 +255,14 @@ def parse_jupyter_message(
 
     if values["msg_type"] not in _KNOWN_MESSAGE_TYPES:
         return UnknownJupyterMessage.model_validate({**values, "raw": raw})
-    return _KNOWN_MESSAGE_ADAPTER.validate_python(values)
+    try:
+        return _KNOWN_MESSAGE_ADAPTER.validate_python(values)
+    except ValidationError:
+        # Known msg_type but content or channel doesn't match the strict
+        # model (e.g., extra execution_state, wrong stdout/stderr name,
+        # or msg delivered on an unexpected channel). Degrade losslessly
+        # instead of aborting the whole call.
+        return UnknownJupyterMessage.model_validate({**values, "raw": raw})
 
 
 # Existing callers use this name for the transport-level message. Keep it as