fix(build): resolve Windows CUDA compiler fallback and clarify runtime ownership

[lusoris]

Summary

Windows CUDA configuration failed with Unknown variable cl_path when vswhere was unavailable or returned no compiler, even if cl.exe was on PATH. Assign the fallback path once and use it for both NVCC and MSVC header discovery. A Meson configure regression exercises the shipped discovery block and runs in the POSIX fast suite.

Also correct native-versus-Python ONNX Runtime ownership documentation: the native CPU archive does not provide CUDA or ROCm execution providers. Runtime pins are unchanged.

This is a held follow-up to #1396. The preceding hygiene PR #1420 was merged into that parent by the old local train, not into master. The parent and this follow-up remain drafts pending full integration acceptance; stacked PRs must not be auto-promoted or merged by the train.

Validation

• Four configure cases pass with Meson 1.10.1 and 1.12.0; both fallback cases fail against the original source with the undefined-variable error.
• Actual Meson --suite=fast registration executes the fixture successfully (1/1 registered test, four internal cases).
• Ruff lint/format and normal pre-commit/commit-message gates pass. Generated changelog freshness is checked.
• The native Windows CI sample linked in Netflix/vmaf#1472 proves compile/link/install on its recorded earlier revision. This configure regression does not establish Windows GPU runtime or numerical parity.
• Full combined make lint, make test, fresh image, golden-data and hosted acceptance remain pending. No merge-readiness claim is made from focused checks.

Bug-status hygiene

• docs/state.md records the Windows fallback bug and the ORT ownership correction.
• Netflix golden-score assertions are unchanged.

Deep-dive deliverables

• Research digest — no digest needed: trivial compiler-path assignment and documentation correction, with executable negative controls.
• Decision matrix — no alternatives: only-one-way fix; both consumers require the compiler path selected by the existing fallback.
• AGENTS.md invariant note — core/src/AGENTS.md records the shared compiler-path contract.
• Reproducer / smoke-test command — below.
• CHANGELOG fragment — changelog.d/fixed/windows-cuda-cl-path-fallback.md and changelog.d/fixed/ort-runtime-owner-comments.md.
• Rebase note — docs/rebase-notes.md records both contracts.

Reproducer

python3 core/test/test_windows_cuda_compiler_discovery.py -v

Native compilation and hardware scoring remain separate validation steps.

Verified exact-source validation update — 2026-09-08

This supersedes the earlier statement that make test remained pending. Current PR head cc022245079f1b7da4005a39bcefc759e2d6ef3d (tree c2d83fb0798f4989874ef550e60de31f69a832a8) passed the configured CPU make test command: exit 0, 134/134 Meson tests passed, with ASM and default LTO enabled. The source was mounted read-only in an isolated offline container; the installed application binary was not used.

Full make -k lint exited 2 because clang-tidy, black, shellcheck, npx, and gosec were missing. cppcheck was also absent, but lint-c stopped before reaching it. Ruff's configured scope and generated-document consistency checks passed; that does not make full lint green. Exact commands, logs, configuration, hashes, and limitations are retained in .workingdir2/evidence/2026-09-08-rc1-full-gates-cc022245/README.md. The commands select verified Meson/Ninja paths and suppress only their offline bootstrap attempts; source compilation, tests, and analyzer failures are preserved.

The Makefile's separate Python golden-data and sanitizer targets were not run. CUDA, SYCL, HIP, Metal, DNN, MCP, and docs generation were disabled. This GCC 15 CPU result does not establish backend, native Windows/macOS, canonical GCC 14 CI, or RC1 acceptance. Full lint and the remaining integration gates are still required; this PR remains draft and held.

Ancestry is verified: cc02224 → a49d54f → current draft parent #1396 at 851d53dc8112e103e5f55544c09b3dc78f02f7da. Thus the CPU receipt applies to this current combined source without a new rebase. The original #1420 merge event produced 834f84063a4640e72b0916a9a5b3a566b6810fec on that parent branch only; the legacy train subsequently rebased it onto master 78c9d2bf at 18:17:41 +0200, before suspension. The hygiene stack was not merged into master.

Related held work: #1422 contains the queue/backpressure repair; #1423 contains guarded train control and its paused migration helper. #1416 configuration convergence remained pending push at this update (remote head d71e4336). Runtime installation belongs to the coordinating operator; this documentation update changes no holds, draft states, or processes.

Canonical CPU golden follow-up — 2026-09-08, exact cc02224

The unchanged original QualityRunnerTest::test_run_vmaf_runner and
QualityRunnerTest::test_run_vmaf_runner_checkerboard nodes both passed
(2/2). Together they cover all three canonical pairs: normal video and
both checkerboard shifts, plus self-reference and feature-score checks.
No original assertion or tolerance changed. Only executable/resource
paths were bound to the exact retained CPU binary and input data; the same
cc02224 source was mounted read-only with a fresh output workspace.

The bounded CPU container run used NumPy distribution/module 2.5.3 and
the hash-verified official PyWavelets 1.10.0 distribution, installed only
in a private read-only overlay. The PyWavelets wheel's packaged module
reports 1.8.0, despite its distribution METADATA reporting 1.10.0. The
two runtime floors are therefore resolved at distribution metadata level;
PyWavelets runtime version identity remains inconsistent. This limitation
is recorded without changing the package, source or golden assertions.

The central local receipt
is retained on the workstation at
.workingdir2/evidence/2026-09-08-rc1-canonical-golden-cc022245/floor-overlay/README.md
(not a hosted GitHub artifact). It includes official package metadata,
wheel/overlay/source/binary/data hashes, exact command and versions, the
initial version-identity preflight refusal, final passing JUnit/logs, and
unchanged-input verification. The initial passing run with existing
dependencies is retained separately in the parent evidence directory.

This supersedes only the earlier pending canonical three-pair CPU golden
status. The complete five-module make test-netflix-golden target, full lint,
sanitisers, other backends/platforms and hosted integration acceptance remain
separate and unverified by this run. This PR remains draft and held; no RC1
readiness or merge approval is claimed.

[lusoris]

Superseded by the pre-rc.1 fixing train in #1425.

Verified by comparing this branch's tree against the train for every path its commits touch: zero differences. Its own commits are on the train already, and everything else it carried was the base-image chain that #1396's squash put on master on 2026-09-15.

Closing is safe; nothing here is lost. If you would rather keep it open until #1425 merges, that works too — it just has nothing left to contribute.

[lusoris]

Superseded by the pre-rc.1 fixing train in #1425.

Closing because the ADR Collision Guard is a required check on #1425 and fails while two open PRs claim the same ADR numbers. Verified before closing:

• Five of the six ADRs this branch adds are byte-identical in fix(rc1): integrate configuration, hooks and runtime repairs #1425.
• The sixth, ADR-1231, differs only because the train's copy has seven extra lines and none missing — a 2026-09-08 verification-update section this branch predates.
• The branch's own commits are on the train. Everything else it carried was the base-image chain that build(docker): define every container base image in one config file (ADR-1231) #1396's squash put on master on 2026-09-15.

The branch is untouched, so this is reversible.