[Fusion] Fix redundant copying of typed input lists

[tobias-tengler]

Ordinary successful warm mutations with typed list variables validate items into temporary storage, then copy them into the final owned array. This repeats a copy and pooled-buffer bookkeeping even though the JSON array length is already known.

This PR fills the final owned array directly while preserving item validation order, first-error paths, nested lists and scalar-to-list fallback. Empty lists retain shared empty storage, and request values remain independently owned.

Method	Version	Mean	Ratio	Allocated	Alloc Ratio
Coerce one product ID (changing value)	main	280.62 ns	1.00	528 B	1.00
Coerce one product ID (changing value)	PR	241.96 ns	0.86	528 B	1.00
Coerce three product IDs (changing values)	main	362.74 ns	1.00	672 B	1.00
Coerce three product IDs (changing values)	PR	320.85 ns	0.88	672 B	1.00
Coerce five product IDs (catalog control)	main	448.75 ns	1.00	816 B	1.00
Coerce five product IDs (catalog control)	PR	401.04 ns	0.89	816 B	1.00
Coerce changing one/three product IDs	main	321.15 ns	1.00	600 B	1.00
Coerce changing one/three product IDs	PR	286.42 ns	0.89	600 B	1.00
Coerce empty list control	main	198.80 ns	1.00	368 B	1.00
Coerce empty list control	PR	180.29 ns	0.91	368 B	1.00
Coerce Int variable control	main	210.43 ns	1.00	416 B	1.00
Coerce Int variable control	PR	196.28 ns	0.93	416 B	1.00
Reject invalid second item control	main	720.23 ns	1.00	2520 B	1.00
Reject invalid second item control	PR	667.31 ns	0.93	2568 B	1.02
Use missing variable default control	main	136.04 ns	1.00	336 B	1.00
Use missing variable default control	PR	136.38 ns	1.00	336 B	1.00
Coerce scalar to list control	main	250.36 ns	1.00	528 B	1.00
Coerce scalar to list control	PR	236.91 ns	0.95	528 B	1.00

The complete production variable-coercion helper, including UTF8 output consumption and owned-output escape, improved 10.8–13.8% for the ordinary one/three-item cases, meeting the localized 5% gate; five items are a derived catalog control. Means include all 360 observations and 36 native MemoryDiagnoser rows from two independent launches per arm (BDN 0.15.8, Release .NET 10.0.12, ARM64, server GC, two CPUs). main is 284f0b5f44cf, measured source c66f6ae4f182, and published source 9f0dee621aeb uses equivalent declaration syntax with identical .NET 10 method bodies; Debug .NET 11 style validation and 34 Release tests passed. Successful allocation is unchanged. The error control adds 48 B (1.9%) for its abandoned three-slot array. Scalar/singleton controls also run faster; no control adjustment or observations were removed. This measures CPU-bound helper elapsed time, not whole-gateway performance.

[github-actions]

Patch coverage

100.0% of changed lines covered (22/22)

File	Covered	Changed	Patch %
…/src/Fusion.Execution/Execution/JsonVariableCoercion.cs	22	22	100.0% 🟢

---

Project coverage: 59.0% (311546/527834 lines)