[SandroMaglione]

Games can now run in a native C11 application embedding QuickJS-ng and SDL3, with game-owned 2D drawing and the existing TypeScript simulation, input and fixed-loop contracts. This is the working Linux MVP from #47, with numeric C kernel compilation and mobile project tooling.

• Add @typeonce/ritmo-native: doctor, build, dev, check, retained rectangles/PNG sprites/debug text, structural keyboard/pointer/touch adapters, lifecycle and safe-area handling.
• Add private, versioned core numeric columns, query views and transaction journals. A TypeScript 6 analyzer emits deterministic C11 for supported numeric systems, retains interpreted fallbacks, reports unsupported constructs and enforces requireCompiled.
• Add streaming per-frame Node/native oracle checks with kernels on and off, 10,000 numeric fuzz cases, C numeric/journal tests, rendering goldens and native benchmarks.
• Add Android Gradle/SDLActivity and iOS Xcode project generators and build/oracle launchers, plus desktop/mobile/sanitizer CI jobs.
• Add the top-down native entry and touch controls, and an external inkline-town integration script using its real simulation with a native 2D town view. Its Three.js renderer is not ported.
• Pin and license QuickJS-ng 0.12.0, SDL 3.2.28 and stb_image.

Local Linux verification: pnpm check (446 runtime tests plus type tests), pnpm pack:check, full core suite in numeric-column mode (272 tests), CTest and screenshot goldens. Top-down and inkline-town each match Node for 600 frames with kernels on and off. Both the 5,000-entity stress fixture and 10,000 numeric fuzz cases pass with ASan/UBSan. The fuzz run verifies actual C kernel execution.

The user approved normalizing NaN sign/payload bits because V8 and QuickJS differ even with kernels disabled. Every other IEEE-754 bit remains exact, including signed zero.

The final same-machine browser benchmark check passed, with every runtime difference within 20% and type metrics within the committed 5% tolerance. The plain benchmark check flags six cases against the shared Apple M1 baseline on this AMD Linux host; both outputs and the saved pre-change measurements are committed. At 5,000 entities, native numeric movement measured 0.0817 ms/frame versus Node JIT 0.2712 ms/frame; see packages/native/bench/results.md.

This stays a draft: the conservative compiler leaves top-down's map/validation/helper-based movement interpreted, so its planned requireCompiled gate is not met. macOS/Windows execution, actual Android/iOS builds and emulator/simulator checks still need platform CI. Mobile project generation alone has been checked locally. PLAN.md and NATIVE_CHECKPOINT.md retain the exact outstanding gates and reproduction commands.

All work remains in this single implementation PR. Do not merge this PR or #47 into main.