This paper presents flutter_kiwi_nlp, a production-oriented Flutter plugin for Korean morphological analysis built on Kiwi. The package exposes one stable Dart API while internally operating two runtime stacks: Native (Dart FFI + C bridge + Kiwi shared library) and Web (Dart JS interop + kiwi-nlp WASM). This design enables a single application codebase across Android, iOS, macOS, Linux, Windows, and Web. Unlike ONNX-export deployment paths, this integration does not require adding an extra generic inference runtime layer because Kiwi already provides the analysis engine and model format used here. The implementation is explicitly aligned with on-device AI requirements: local inference execution, reduced text egress by default, predictable latency without per-request network dependence, and operational fallback controls for model provisioning. Empirically, repeated benchmark trials show a warm-path throughput reversal (steady-state path, initialization excluded): Flutter warm throughput can exceed the Python-native baseline: 3.12x on desktop, 3.87x on iOS simulator (debug), and 1.52x on Android emulator (release). Added Linux/Windows CI rows show clear execution-mode sensitivity: single mode is 0.64x of Python throughput on both OSes, while batch mode is 1.54-1.55x faster than Python on the same runners. The updated benchmark pipeline also adds boundary-decomposed measurements (pure processing vs full JSON path), quantifying serialization/parsing overhead at 13.89-47.93% depending on runtime profile. This warm-path reversal does not imply end-to-end session superiority: cold-start initialization and short-session effective-throughput remain weaker than the Python baseline on all measured platforms. Cross-runtime linguistic agreement remains close on gold corpora (88.39% vs 88.58% token agreement; 84.90% vs 85.55% POS agreement). On native targets, inference can run fully offline after one-time model provisioning. This paper provides an implementation-complete, reproducible specification of API contract, runtime parity rules, build automation hooks, security boundaries, failure taxonomy, benchmark protocol, and quantified ecosystem survey signals.