The delegate-conversion tests all miss the instance-keyed cache

[lahma]

#3521 keyed both delegate caches by the target type. HostDelegateConversionTests pins it through three shapes — two engines on one Prepared<Script>, the same in the other order, and the two alternating — and every one of them gives each engine a function of its own.

That means all three exercise the AST-keyed cache. None exercises the instance-keyed one: no test asks a single function instance for two different delegate types, which is the entire reason the outer ConditionalWeakTable<object, TypeKeyedCache<Delegate>> carries a TypeKeyedCache rather than a bare Delegate.

The gap, measured

Keying only the instance lane on a constant — so it behaves as it did before #3521 while the AST lane stays keyed by target type:

-if (!boundDelegates.TryGetValue(type, out var bound))
+if (!boundDelegates.TryGetValue(typeof(object), out var bound))
...
-bound = boundDelegates.GetOrAdd(type, targetBinder(functionInstance)!);
+bound = boundDelegates.GetOrAdd(typeof(object), targetBinder(functionInstance)!);

Passed!  - Failed: 0, Passed: 3

All three existing tests stay green. A simplification that folded the instance lane back to a plain weak table would have landed unnoticed.

The test added here fails under exactly that mutation, with the defect's own message:

Failed OneEngineConvertingOneFunctionToTwoDelegateTypes
  System.ArgumentException : Object of type 'Notify' cannot be converted to type 'Transform'.

and passes on main as it stands. It also converts back to the first type afterwards, so it says the two entries coexist rather than evict each other.

Test-only; no production change. Found while auditing process-wide caches for #3426 — I had an equivalent local fix and test predating #3521, and this is the only part of it main did not already have.

Verification

Jint.Tests.PublicInterface HostDelegateConversionTests: 4/0 on net472, net8.0 and net10.0.

🤖 Generated with Claude Code

https://claude.ai/code/session_014W5mbjGhyvgAS4pivXoc4S