[tobias-tengler]

Variable coercion allocated several objects for every field of an input object variable: JsonVariableCoercion read JsonProperty.Name up to three times per field (field lookup, error path segment and the coerced ObjectFieldNode), and each read decodes a new UTF-16 string. Every coerced field also got a new NameNode.

This PR looks input fields up by their raw UTF-8 property name. FusionFieldDefinitionCollection<T> gets a TryGetField(ReadOnlySpan<byte>) overload: collections with up to eight accessible fields are scanned with Ascii.Equals, larger ones widen the name into a stack buffer and use the ReadOnlySpan<char> alternate lookup of the existing frozen dictionary. FusionEnumValueCollection does the same and drops its sorted byte index, so no collection keeps a second copy of its names. The coerced ObjectFieldNode reuses a NameNode cached on FusionInputFieldDefinition, and the error path uses the schema's field name. JSON allows escaping any character of a property name ({"\u0069d": 1} is a valid way to send id), so a property name containing an escape sequence falls back to the unescaped lookup, only when the UTF-8 lookup misses.

JsonMarshal.GetRawUtf8PropertyName exists since .NET 10, so net8.0 and net9.0 keep the string lookup. Per schema this adds the cached NameNode (40 B per input field definition) and one lookup struct per collection (16 B on net9.0 and later), nothing else.

Benchmarks

VariableCoercionBenchmark (added in this PR), net11 in-process.

Small input (2 objects, 5 fields)

Method	Version	Mean	Ratio	Allocated	Alloc Ratio
Variable coercion	main	737.2 ns	1.00	1.67 KB	1.00
Variable coercion	PR	555.8 ns	0.75	0.98 KB	0.59

Large input with a list of 20 input objects (22 objects, 66 fields)

Method	Version	Mean	Ratio	Allocated	Alloc Ratio
Variable coercion	main	8.262 μs	1.00	18.27 KB	1.00
Variable coercion	PR	5.696 μs	0.69	9.48 KB	0.52

Variable coercion allocates 41% less for the small input and 48% less for the large input, and is 25% and 31% faster.

[Copilot]

🟡 Changes recommended

The enum refactor introduces an avoidable array allocation for every enum collection.

1 open finding

• Avoid unnecessary parallel array allocation in Utf8NameIndex.Create · New

What changed in this PR

Optimizes Fusion input-object variable coercion by matching raw UTF-8 property names and reusing cached syntax nodes.

Changes:

• Adds a shared UTF-8 name index.
• Uses UTF-8 lookup with escaped-name fallback.
• Adds lookup and coercion coverage.

File	Description
VariableCoercionHelperTests.cs	Tests escaped JSON field names.
FusionInputFieldDefinitionCollectionTests.cs	Tests UTF-8 field lookup.
JsonVariableCoercion.cs	Uses UTF-8 lookup and cached names.
FusionInputFieldDefinition.cs	Caches each field’s NameNode.
Utf8NameIndex.cs	Implements shared UTF-8 indexing.
FusionInputFieldDefinitionCollection.cs	Adds span-based field lookup.
FusionEnumValueCollection.cs	Adopts the shared index.

🧠 Review effort: Balanced

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[github-actions]

Patch coverage

100.0% of changed lines covered (47/47)

File	Covered	Changed	Patch %
…/Collections/FusionEnumValueCollection.cs	8	8	100.0% 🟢
…/Collections/FusionFieldDefinitionCollection.cs	23	23	100.0% 🟢
…/src/Fusion.Execution.Types/FusionInputFieldDefinition.cs	2	2	100.0% 🟢
…/src/Fusion.Execution/Execution/JsonVariableCoercion.cs	14	14	100.0% 🟢

---

Project coverage: 58.7% (315520/537948 lines)