_sanitize leaks IntEnum/StrEnum members onto the settings wire, breaking subscribers without the defining package

[cboulay]

Summary

ezmsg.core.settingsmeta._sanitize checks isinstance(value, (bool, int, float, str)) before it checks isinstance(value, enum.Enum). IntEnum and StrEnum members are instances of int/str, so they short-circuit on the first branch and are returned verbatim, un-sanitized.

Those values then travel to the graph server and out over the settings event / snapshot / metadata wires as pickled enum members. Any subscriber that does not have the defining package installed cannot pickle.loads them, and the whole subscription dies with ModuleNotFoundError.

Where

ezmsg/core/settingsmeta.py (v3.9.0, and current dev):

def _sanitize(value: Any) -> Any:
    if value is None or isinstance(value, (bool, int, float, str)):
        return value                     # <-- IntEnum / StrEnum returned as-is
    if isinstance(value, enum.Enum):
        return _sanitize(value.value)    # <-- unreachable for IntEnum / StrEnum
    ...

Reproduction

import enum
import ezmsg.core as ez
from ezmsg.core.settingsmeta import settings_structured_value, settings_schema_from_type

class Rate(enum.IntEnum):
    FAST = 5

class Mode(enum.StrEnum):
    A = "a"

class Plain(enum.Enum):
    X = "x"

class S(ez.Settings):
    rate: Rate = Rate.FAST
    mode: Mode = Mode.A
    plain: Plain = Plain.X

print(settings_structured_value(S()))
# {'rate': <Rate.FAST: 5>, 'mode': <Mode.A: 'a'>, 'plain': 'x'}
#          ^^^^^^^^^^^^^^^  ^^^^^^^^^^^^^^^^^^^ leaked          ^^^ correctly sanitized

print([(f.name, f.default) for f in settings_schema_from_type(S).fields])
# [('rate', <Rate.FAST: 5>), ('mode', <Mode.A: 'a'>), ('plain', 'x')]

Affected wires

_sanitize is the sanitizer for everything settings-shaped that leaves the process, so the leak reaches all of these:

• SettingsSnapshotValue.structured_value and .repr_value — backendprocess.py::_settings_snapshot_value
• SettingsFieldMetadata.default and .choices — settings_schema_from_type, hence ComponentMetadata.settings_schema in GraphMetadata
• and therefore SettingsChangedEvent on GraphContext.subscribe_settings_events(), GraphContext.settings_snapshot(), and the component metadata a client reads back from the graph server

Observed behaviour

Our experiment-control UI runs in its own venv and deliberately does not depend on the pipelines' acquisition stack. Its settings-event forwarder died on every reconnect:

File "ezmsg/core/graphcontext.py", line 581, in _subscribe_pickled_stream
    value = pickle.loads(payload)
ModuleNotFoundError: No module named 'pycbsdk'

The setting in question is ezmsg.blackrock.CereLinkSignalSettings.subscribe_rate: SampleRate, where pycbsdk.session.SampleRate is an IntEnum.

Two things make this worse than a single dropped message:

1. It wedges permanently. The subscriber reconnects with after_seq=0, which replays the graph server's retained event history — including the event it cannot decode. So one un-decodable retained event kills settings observation for the whole life of the graph server, not for one message. _subscribe_pickled_stream has no per-payload recovery: a failed pickle.loads propagates out of the generator and tears down the subscription.
2. The information was never needed. _sanitize's entire job is to reduce settings to JSON-safe scalars so that observers need none of the graph's dependencies. Reordering the two checks yields {'rate': 5, 'mode': 'a', 'plain': 'x'} — a pycbsdk-free payload with no loss of anything a consumer uses (json.dumps already renders an IntEnum as its integer, so rendered output is unchanged).

Suggested fix

Move the enum.Enum check above the primitives check:

def _sanitize(value: Any) -> Any:
    if isinstance(value, enum.Enum):
        return _sanitize(value.value)
    if value is None or isinstance(value, (bool, int, float, str)):
        return value
    ...

This is generic — it covers IntEnum, StrEnum, IntFlag, and any other mixin-enum from any dependency, not just this one.

Happy to open a PR with the change plus a regression test asserting _sanitize leaves no non-builtin types in its output for an IntEnum/StrEnum-valued Settings.

Possibly worth considering separately

• _subscribe_pickled_stream could survive a bad payload — log and skip rather than terminate the subscription. Even with the sanitizer fixed, the wire is raw pickle, so the next accidental leak reproduces the same permanent wedge in a client that is doing nothing wrong.
• These wires arguably need not be pickle at all. After this fix, SettingsChangedEvent / TopologyChangedEvent / ProfilingTraceStreamBatch contain only JSON-representable data plus UUID, enum, tuple and one opaque bytes field (SettingsSnapshotValue.serialized, which nothing in the tree ever unpickles — ezmsg-dashboard only reports serialized is not None). A JSON envelope would make the cross-venv contract structural instead of conventional. Related: GraphServer pickles snapshots under the global command lock, stalling a starting graph #257.