feat(minimax-h3): add native Windows RTX path

[yifeif-nv]

Background

The native runtime and build flow relied on POSIX shared-library and process behavior, while MiniMax H3 could retain several very large plans at once. This change adds a focused native Windows path for the fixed MiniMax H3 video profile without broadening Windows support to unrelated runtime modes.

Exit Criteria

• Build the CLI, core runtime, TensorRT-RTX backend, and MiniMax H3 model DLL from explicit local CUDA 12.9 and TensorRT-RTX SDK roots.
• Build the six H3 plans in isolated processes and stream them into a staged bundle without materializing every plan in host memory.
• Verify each plan before file-backed deserialization, prebind shared device buffers, and prevent denoiser/VAE residency overlap.
• Keep the existing standard TensorRT path unchanged and avoid publishing checkpoints, SDK files, generated media, or machine-identifying provenance.

Implementation

• Adds a small cross-platform shared-library boundary, portable CLI child-process/temp-path handling, model-only CMake selection, and an explicit Windows SDK/output layout.
• Adds one model-owned staged builder with a 16 GiB tactic-workspace limit, a 32 GiB resident-weight budget, direct-to-file TensorRT serialization, content-bound resume, and sanitized bundle metadata.
• Streams validated bundle ranges into TensorRT-RTX, applies the weight budget before context creation, prebinds H3 split-plan/VAE buffers, and releases each runtime stage before loading the next.
• Preserves BF16 attention range instead of narrowing H3 Q/K/V tensors to FP16.
• Documents the narrow local workflow and its authorization, licensing, output, and validation boundaries; no downloader is included.

Validation

• ruff check python/tensorrt_model_connect/bundle_writer.py python/tensorrt_model_connect/engine_builder.py python/tensorrt_model_connect/trt_compat.py python/tensorrt_model_connect/families/minimax_h3: passed.
• python3 -m pytest -q python/tensorrt_model_connect/families/minimax_h3/tests/test_staged_build.py python/tensorrt_model_connect/families/minimax_h3/tests/test_trt_builders.py tests/builder/test_bundle_writer.py tests/builder/test_trt_compat_boundary.py tests/builder/test_max_batch_size_cli.py tests/builder/test_engine_builder_extended.py: 59 passed, 1 skipped.
• python3 -m pytest -q tests/tools/test_check_doc_file_references.py tests/tools/test_model_plugin_encapsulation_static.py tests/tools/test_test_impact.py tests/tools/test_coverage_map.py: 476 passed.
• python3 -m tools.community_ci impact --base github/main: passed; the Windows helper selects only minimax-h3-768p, builder units, and the native rebuild lane.
• python3 -m tools.community_ci source-quality --base github/main: passed, including max CCN 10, changed-file lint/formatting, and 158 architecture-contract tests.
• python3 -m tools.community_ci unit --scope all: 3,572 passed, 2 skipped.
• Linux focused CMake configure/build of trtmc_platform_cpp_tests, trtmc_model_minimax_h3, and the H3 external-prebinding test: passed.
• ctest --test-dir <build> --output-on-failure -L platform: 40/40 passed.
• ctest --test-dir <build> --output-on-failure -R '^test_minimax_h3_trt_module_bind_external$': 1/1 passed.
• Forced compilation with the RTX delete/alias/communicator API-age overrides enabled: passed against TensorRT 11 headers.
• python3 tools/legal_headers.py --check: 5,902 tracked files, 0 findings.
• PowerShell/MSVC compilation, Windows DLL loading, TensorRT-RTX execution, and real MiniMax H3 build/inference were not run locally because this environment has no Windows toolchain, TensorRT-RTX SDK, or authorized H3 checkpoint.

Notes For Future Readers

• Review the platform/CMake boundary first, the Python staged builder second, and the RTX/H3 runtime lifetime changes last.
• This supports one fixed SM 120, BF16, batch-1, single-device, decoded-video profile. It does not add Windows serving, distributed execution, audio output, or general model-family support.
• Plans and generated bundles are local, environment-bound artifacts. Rebuild with authorized inputs when the checkpoint or SDK/backend cohort changes.