fix: isolate type decoders from shared signature annotation state

[maksimzayats]

Description

This removes the shared _decoder mutation from signature parsing.

Previously, SignatureModel.create() could attach a decoder directly to the annotated type, which made decoder behavior process-global. That meant one handler, app, or DI provider could accidentally affect another if they used the same annotation class.

Now, signature parsing no longer mutates user-defined or stdlib classes. Request parsing stays scoped to the current handler or provider through the existing default_deserializer(type_decoders=...) path.

class UserId:
    def __init__(self, value: str) -> None:
        self.value = value


@get("/decoded", type_decoders=[(lambda t: t is UserId, lambda t, v: t(f"decoded:{v}"))])
def decoded(user_id: UserId) -> str:
    return user_id.value


@get("/plain")
def plain(user_id: UserId) -> str:
    return user_id.value

Old behavior:

# Building /decoded mutated the shared UserId type:
UserId._decoder = lambda t, v: t(f"decoded:{v}")

# So /plain could start decoding too, even without its own decoder.
# /decoded?user_id=1 -> "decoded:1"
# /plain?user_id=1   -> "decoded:1"   # leaked

New behavior:

# No mutation on UserId.
# Decoding stays scoped to the current handler.

# /decoded?user_id=1 -> "decoded:1"
# /plain?user_id=1   -> 400 validation error

---

📚 Documentation preview 📚: https://litestar-org.github.io/litestar-docs-preview/4639

[sobolevn]

I would categorize this as a tricky optimization instead of just a bug. Current behavior looks more correct to me. However, we also have to measure the perf impact of this change.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 97.87%. Comparing base (9587ba1) to head (9ab8b99).
⚠️ Report is 13 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #4639      +/-   ##
==========================================
- Coverage   97.87%   97.87%   -0.01%     
==========================================
  Files         292      292              
  Lines       15077    15067      -10     
  Branches     1703     1701       -2     
==========================================
- Hits        14757    14747      -10     
  Misses        184      184              
  Partials      136      136              

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.

[maksimzayats]

Hi @sobolevn ! Hmm, the old behavior looks quite strange to me tbh... and it looks like more a bug, but I'd love to hear other maintainers opinion on this

[sobolevn]

Context: jcrist/msgspec#497 and jcrist/msgspec#483

[provinzkraut]

Agree with @sobolevn. Yes, it's a bit iffy to store things globally like this, but it's an optimisation we deemed worth it, as the lookup process is quite expensive. If you want to remove this, I'd need comprehensive benchmarks to show there's no negative impact on performance