…dens matrix dimensions

Issue OPCFoundation#3546 reported that the libfuzz CI pipeline (which feeds a zip of
corpus inputs sequentially) caught a DataTypeNode/VariableNode round-trip
failure that the local NUnit `EncoderTests` runner did not. The local
miss had TWO root causes:

  1. `Fuzz.Tests/FuzzTargetTestsBase.FuzzCrashAssets` silently swallowed
     every exception — counter was incremented but never asserted. So a
     residual `Assets/crash-*` could sit in the tree without breaking
     `dotnet test`.
  2. The single crash asset on disk reproduces an `ArgumentException` from
     `MatrixOf<T>(values, dimensions)` (deliberate security validation
     for overflow / negative / length-mismatch dimensions) escaping
     `BinaryDecoder.ReadVariantValue` and `ReadEncodeableMatrix<T>`. The
     fuzz target only swallows `ServiceResultException(BadDecodingError /
     BadEncodingLimitsExceeded)`, so the unhandled `ArgumentException`
     surfaced as a "crash" in the pipeline.

Both are fixed here.

### Production fix
`Stack/Opc.Ua.Types/Encoders/BinaryDecoder.cs`:
  * Wrap the variant-matrix construction switch in `ReadVariantValue`
    (covers all 24 BuiltInType cases) in a try/catch that converts
    `ArgumentException` from `MatrixOf<T>` validation into a
    `ServiceResultException(BadDecodingError)`. Attacker-controlled wire
    dimensions are now rejected through the standard decoder channel.
  * Same treatment for both `ReadEncodeableMatrix<T>` overloads which
    call `ArrayOf<T>.ToMatrix(dimensions)`.

The fix preserves the deliberate `MatrixOf<T>` ctor validation (the
security hardening at `MatrixOf.cs:178-217`); it only routes that
rejection through the same channel as the rest of the decoder.

### Regression coverage
`Tests/Opc.Ua.Types.Tests/Encoders/BinaryDecoderTests.cs`:
  * `ReadMatrixWithOverflowingDimensionsThrowsBadDecodingError`
  * `ReadMatrixWithNegativeDimensionThrowsBadDecodingError`
  * `ReadMatrixWithLengthMismatchThrowsBadDecodingError`

`Fuzzing/Opc.Ua.Encoders.Fuzz.Tests/Issue3546RoundTripTests.cs`:
  * 8 synthetic round-trip tests exercising populated `DataTypeNode` and
    `VariableNode` through `FuzzableCode.FuzzBinaryEncoderIndempotentCore`
    (the exact 3rd-gen IsEqual gate that originally failed in CI). All
    pass, confirming that the modern `Equals` overhaul (`ArrayOf<T>`,
    `ExtensionObject.Equals`, IEquatable on `ReferenceNode` /
    `RolePermissionType`) has implicitly resolved the IsEqual
    non-determinism that OPCFoundation#3546 was filed for.

### Fuzz harness
`Fuzzing/Common/Fuzz.Tests/FuzzTargetTestsBase.cs`:
  * `FuzzCrashAssets` now collects every per-asset failure and finishes
    with `Assert.That(failures, Is.Empty, ...)`. CI breaks loud when a
    residual crash asset reproduces — the exact gap that originally let
    OPCFoundation#3546 slip past `dotnet test`.

`.github/agents/fuzz-tester.agent.md`: rewrote the "does not currently
Assert" note to reflect the new assertive behaviour and link the gap
closure to OPCFoundation#3546.

### Permanent corpus seeds
`Fuzzing/Opc.Ua.Encoders.Fuzz.Tools/Encoders.Testcases.cs`: added
`DataTypeNodeMessage` and `VariableNodeMessage` message-encoder methods
that populate the fields originally implicated (non-empty References,
RolePermissions, DataTypeDefinition with an `ExtensionObject` body for
`DataTypeNode`; Value `Variant` carrying an `ExtensionObject`,
non-empty ArrayDimensions/References/RolePermissions for `VariableNode`).

`Fuzzing/Opc.Ua.Encoders.Fuzz.Corpus/Testcases.{Binary,Json,Xml}/`:
6 new seed files (`datatypenodemessage.{bin,json,xml}` and
`variablenodemessage.{bin,json,xml}`) generated by the tool. These flow
automatically into `FuzzGoodTestcases` so the round-trip is exercised
through every fuzz target on every test run.

### Verification (net10.0 + net48)
* Encoders.Fuzz.Tests net10.0: 4625 / 0 (baseline 4131 + 6 new seeds *
  ~81 targets effectively across the matrix + 8 synthetic + 3 other
  new tests)
* Encoders.Fuzz.Tests net48:   4625 / 0
* Types.Tests (BinaryDecoder/Matrix/Variant/EncodeableMatrixRoundTrip
  subset) net10.0: 985 / 0 (baseline 982 + 3 new regression tests).

The single pre-existing `Assets/crash-0483107a...` no longer escapes any
fuzz target — the BinaryDecoder fix above turns it into a clean
parser-rejected input.