Push adapter initialization does not forward annotation arguments correctly

[AdamGlustein]

Push adapters (potentially all adapters, haven't checked) are normalizing annotation arguments even if they are not within a csp.ts object.

Describe the bug

To Reproduce

from datetime import datetime, timedelta
from typing import Dict, TypeVar

import csp
from csp.impl.pushadapter import PushInputAdapter
from csp.impl.wiring import py_push_adapter_def

T = TypeVar("T")


class _AdapterImpl(PushInputAdapter):
    def __init__(self, ts_typ: T):  # noqa
        print(ts_typ)
        raise

    def start(self, starttime: datetime, endtime: datetime):
        pass

    def stop(self):
        pass

MyPushAdapter = py_push_adapter_def(
    "adapter_def",
    _AdapterImpl,
    csp.ts[T],
    ts_typ=T,
)

@csp.graph
def my_graph() -> csp.ts[Dict[str, int]]:
    data = MyPushAdapter(ts_typ=Dict[str,int])
    return data

csp.run(my_graph, starttime=datetime.utcnow(), endtime=timedelta())

Prints: class <dict> instead of typing.Dict[str, int]

Expected behavior

Should print typing.Dict[str, int]

Error Message

Runtime Environment

0.11.1
3.11.11 (main, Jan 30 2025, 14:30:53) [GCC 13.3.0]
linux

Additional context

[robambalu]

This is somewhat by design since csp normalizes all typing types to basic python types for type mapping at the c++ level when passing type info to c++ adapters. The same mechanism is used regardless of the adapter impl ( python or c++ ) What practical issue is this causing?

[AdamGlustein]

This is somewhat by design since csp normalizes all typing types to basic python types for type mapping at the c++ level when passing type info to c++ adapters. The same mechanism is used regardless of the adapter impl ( python or c++ ) What practical issue is this causing?

The variable ts_typ here is just a regular argument to the __init__ function so it shouldn't be normalized, that is the problem.

@arhamchopra is trying to write an adapter for polling files to output dictionaries of structs and ran into this.

[robambalu]

Again the problem is that the same code is used to pass type info to c++ ( this is used by output adapters regularly ) I think now that typing is much more mature than when we started maybe the best thing to do is actually stop normalizing all together and passing typing constructs down to c++ instead