chore: post-rename + post-vulkan-removal sweep (clang-rt-19, vmaf-tune choices, docs)

.github/workflows/go-ci.yml

@@ -44,8 +44,11 @@ jobs:
           sudo apt-get install -y --no-install-recommends meson ninja-build pkg-config
 
       - name: Build libvmaf (CPU only) for cgo link
+        # Post-ADR-0700: meson.build lives at core/meson.build, so the
+        # source dir must be passed explicitly (or run from core/). Pre-rename
+        # `meson setup core/build-cpu` worked because root had meson.build.
         run: |
-          meson setup core core/build-cpu \
+          meson setup core/build-cpu core \
             -Denable_cuda=false \
             -Denable_sycl=false \
             -Denable_hip=false \

changelog.d/fixed/containerfile-post-rename-sweep.md

@@ -0,0 +1,116 @@
+- **dev/Containerfile**: add `libclang-rt-19-dev` to the build-deps apt layer so
+  that `b_sanitize=address,undefined` meson builds link correctly against the
+  clang-19 ASan/UBSan runtime libraries (previously the linker failed with
+  `cannot find libclang_rt.asan.a`).
+- **vmaf-tune `--score-backend`**: remove `vulkan` from `ALL_BACKENDS` /
+  `DEFAULT_FALLBACKS` in `score_backend.py` and from all argparse `choices=`
+  tuples, following the Vulkan backend removal in ADR-0726. Requesting
+  `--score-backend vulkan` now raises a `ValueError` at startup.
+- **vmaf-tune docs**: update `docs/usage/vmaf-tune-score-backend.md` to reflect
+  the `cuda → sycl → hip → cpu` fallback chain and remove the Vulkan row from
+  the accepted-values table.
+- **Go CI (`.github/workflows/go-ci.yml`)**: post-ADR-0700 sweep miss — the
+  cgo-link build step ran `meson setup core/build-cpu` from repo root, which
+  fails with `Neither source directory 'core/build-cpu' nor build directory
+  None contain a build file meson.build.` since the rename moved `meson.build`
+  into `core/`. Pass the source dir explicitly (`meson setup core/build-cpu
+  core ...`). Restores Go CI green on master.
+- **SYCL integer extractors (`core/src/feature/sycl/integer_adm_sycl.cpp`,
+  `integer_vif_sycl.cpp`)**: add the missing `close_fex_sycl` forward
+  declaration. The init-failure cleanup paths added by ADR-0784 (sycl-init-
+  failure-cleanup-leaks) call `close_fex_sycl(fex)` from within
+  `init_fex_sycl`, but the function is defined later in the file as
+  `static int` without a forward decl, so SYCL builds fail with
+  `error: use of undeclared identifier 'close_fex_sycl'`. Matches the pattern
+  already used by `float_*_sycl.cpp`. Restores Linux GCC (all backends),
+  Ubuntu SYCL, Ubuntu SYCL + CUDA, and macOS SYCL builds.
+- **CUDA SpEED kernel embed (`core/src/meson.build`)**: register
+  `core/src/feature/cuda/speed/speed_score.cu` in the `cuda_cu_sources`
+  dict so it is compiled to PTX and embedded as the `speed_score_ptx`
+  C string. Both `speed_chroma_cuda.c` and `speed_temporal_cuda.c`
+  declare `extern const char speed_score_ptx[]` and call
+  `cuModuleLoadData(&module, speed_score_ptx, ...)`, but the kernel
+  source was orphaned in the meson dict (added by ADR-0567 but never
+  wired). Without this entry the link fails with `undefined reference
+  to 'speed_score_ptx'`. Restores Docker Image Build, Linux GCC (all
+  backends), Ubuntu CUDA, Ubuntu CUDA Static, Ubuntu SYCL + CUDA, and
+  Windows MSVC + CUDA build legs.
+- **Windows MinGW64 test (`core/test/test_gpu_dispatch_runtime.c`)**: map
+  POSIX `setenv` / `unsetenv` to `_putenv_s` on `_WIN32 && !__CYGWIN__`.
+  MinGW's libc doesn't expose POSIX env routines so the build failed with
+  `implicit declaration of function 'setenv'`. Restores the Windows
+  MinGW64 (CPU) leg.
+- **Metal kernel coverage audit (`core/test/test_metal_kernel_coverage_audit.c`)**:
+  update the kernel-basename list to use the *registered extractor name*
+  (`motion_v2`, matching `integer_motion_v2_metal.mm`'s actual `.name =
+  "motion_v2_metal"`), not the `.mm` filename. The audit asserted
+  `integer_motion_v2_metal` existed and failed because that name is
+  never registered — the established name is the short form, consistent
+  with `dispatch_strategy.c`, `test_metal_kernel_registration.c`,
+  `test_metal_motion_v2_parity.c`, and `test_metal_smoke.c`. Restores
+  the macOS Clang (CPU + Metal) and Metal (T8-1 scaffold) test legs.
+- **Windows MSVC pthread_once (`core/src/compat/win32/pthread.h`)**:
+  add `pthread_once_t` / `PTHREAD_ONCE_INIT` / `pthread_once()` to the
+  Win32 pthread shim. `iqa/ssim_simd.h` declares
+  `iqa_ssim_install_dispatch_once(pthread_once_t *guard, void (*installer)(void))`
+  and `float_ssim.c` / `float_ms_ssim.c` use a static
+  `pthread_once_t s_dispatch_guard = PTHREAD_ONCE_INIT` for the SIMD
+  dispatch install (TSan race audit 2026-05-30). MSVC fails the whole
+  Windows + CUDA leg because the existing win32 shim only provides
+  mutex / cond / thread, not once-init. Backs onto
+  `InitOnceExecuteOnce` (Windows Vista+, which is the documented
+  floor for this shim). Restores the MSVC + CUDA and MSVC + oneAPI SYCL
+  build-only legs.
+- **ASan/TSan: `test_y4m_alloc_failure` (`core/test/test_y4m_alloc_failure.c`)**:
+  skip under sanitizer builds. The test caps virtual address space via
+  `setrlimit(RLIMIT_AS, 256 MiB)` to force the y4m parser's `malloc` to
+  fail; AddressSanitizer / ThreadSanitizer / MemorySanitizer reserve a
+  multi-TB shadow virtual region at startup, so any subsequent ASan
+  internal `mmap` blows past the 256 MiB cap and the test process
+  aborts with "ERROR: Failed to mmap" before the parser even runs. The
+  guarded regression (dst_buf-NULL bug) is C-level, not sanitizer-
+  surfaceable, so skipping under sanitizers gives up no coverage.
+  Restores the `Sanitizers — ASan + UBSan (PR gate)` and `ASan + UBSan
+  + MSan (address / thread)` legs.
+- **MSVC C17 `ATOMIC_VAR_INIT` (`core/src/feature/iqa/ssim_tools.c`)**:
+  use direct value initialisation instead of `ATOMIC_VAR_INIT(NULL)`. The
+  macro was deprecated in C17 and MSVC's `<stdatomic.h>` does not provide
+  it (`error C2099: initializer is not a constant`). Direct
+  initialisation `static _Atomic(...) x = NULL` has been semantically
+  equivalent to `ATOMIC_VAR_INIT(NULL)` on every conforming C11
+  implementation. Unblocks the MSVC + CUDA and MSVC + oneAPI SYCL builds.
+- **MinGW64 + MSVC `mkdtemp` (`core/test/test_mkdirp.c`)**: replace
+  the POSIX-only `mkdtemp` call with a portable `vmaf_mkdtemp_portable`
+  helper that picks a temp root from `TMPDIR`/`TEMP`, generates a unique
+  suffix from pid+time+counter, and creates the directory with `mkdir`
+  (POSIX) or `_mkdir` (MSVCRT). MinGW64's libc has `mkdtemp` but the
+  call returns NULL when `/tmp/` isn't where the runtime expects;
+  MSVCRT lacks the function entirely. Restores `test_mkdirp` on the
+  Windows MinGW64 (CPU) leg.
+- **Tiny AI `iter_frames` test fixture (`ai/tests/test_frame_loader.py`)**:
+  `_popen_factory`'s `fake_popen` now accepts the `stderr` keyword that
+  `iter_frames` passes in production (captured for ffmpeg diagnostics
+  surfacing). Pre-fix every `iter_frames(..., popen=_popen_factory(...))`
+  test failed with `TypeError: fake_popen() got an unexpected keyword
+  argument 'stderr'`.
+- **Tiny AI parquet rollback test (`ai/tests/test_parquet_utils.py`)**:
+  monkey-patch `parquet_utils._write_v2` to raise instead of subclass-
+  overriding `DataFrame.to_parquet`. The v2 write path goes through
+  `pq.write_table`, not the DataFrame method, so the old subclass mock
+  was a no-op and `pytest.raises(RuntimeError)` never tripped (`DID NOT
+  RAISE`). The rollback semantics being tested (temp file cleanup on
+  serialiser exception) are now properly exercised.
+- **MCP smoke `_run_benchmark` test (`mcp-server/vmaf-mcp/tests/test_probe_findings_2026_05_17.py`)**:
+  align `test_bug3_run_benchmark_surfaces_silent_pipefail` with ADR-0608 E-1
+  contract (raise `RuntimeError` so MCP marks `isError=True`); test was
+  still asserting the pre-ADR-0608 dict-with-`error` shape.
+- **Tiny AI `_FakeProcess.wait()` (`ai/tests/test_frame_loader.py`)**: accept
+  the `timeout` keyword that `iter_frames` passes per the ADR-0608 progress
+  watchdog.
+- **Tiny AI manifest tests (`ai/tests/test_data_datasets_branches.py`)**:
+  use deterministic 64-char hex digests; pydantic validator on
+  `ManifestEntry.sha256` now rejects the old 8-char shorthand.
+- **Clang-Tidy NOLINTs**: inline justifications on three findings on files
+  touched by this PR — `concurrency-mt-unsafe` (getenv in single-threaded
+  test), `bugprone-reserved-identifier` (`_POSIX_C_SOURCE` feature-test
+  macro), `readability-non-const-parameter` (POSIX pthread_once_t guard).

core/src/compat/win32/pthread.h

@@ -195,4 +195,34 @@ static inline int pthread_cond_broadcast(pthread_cond_t *cond)
     return 0;
 }
 
+/* pthread_once shim — backs onto INIT_ONCE / InitOnceExecuteOnce (Vista+).
+ * Used by ssim_simd.h / float_ssim.c / float_ms_ssim.c to install the
+ * ISA-best dispatch exactly once per process across the libvmaf thread
+ * pool. Spec: the init routine takes no argument and returns void —
+ * matches the (void(*)(void)) callbacks used by iqa_ssim_install_dispatch_once.
+ */
+typedef INIT_ONCE pthread_once_t;
+#define PTHREAD_ONCE_INIT INIT_ONCE_STATIC_INIT
+
+static BOOL CALLBACK vmaf_pthread_once_trampoline(PINIT_ONCE InitOnce, PVOID Parameter,
+                                                  PVOID *Context)
+{
+    (void)InitOnce;
+    (void)Context;
+    void (*init_routine)(void) = (void (*)(void))Parameter;
+    if (init_routine)
+        init_routine();
+    return TRUE;
+}
+
+static inline int pthread_once(pthread_once_t *once_control, void (*init_routine)(void))
+{
+    if (!once_control || !init_routine)
+        return EINVAL;
+    if (!InitOnceExecuteOnce(once_control, vmaf_pthread_once_trampoline,
+                             (PVOID)(uintptr_t)init_routine, NULL))
+        return EINVAL;
+    return 0;
+}
+
 #endif /* VMAF_COMPAT_WIN32_PTHREAD_H_ */

core/src/feature/iqa/ssim_tools.c

@@ -105,6 +105,7 @@ static void iqa_ssim_dispatch_trampoline(void)
         installer();
 }
 
+/* NOLINTNEXTLINE(readability-non-const-parameter) — POSIX pthread_once mutates *guard. */
 void iqa_ssim_install_dispatch_once(pthread_once_t *guard, void (*installer)(void))
 {
     /* `guard` parameter is retained for API symmetry with the

core/src/feature/sycl/integer_adm_sycl.cpp

@@ -1281,6 +1281,7 @@ static void conclude_adm_csf_den(const uint64_t *accum, int h, int w, int scale,
 static void enqueue_adm_work(void *queue_ptr, void *priv, void *shared_ref, void *shared_dis);
 static void adm_pre_graph(void *queue_ptr, void *priv);
 static void adm_post_graph(void *queue_ptr, void *priv);
+static int close_fex_sycl(VmafFeatureExtractor *fex); /* forward decl for init error paths */
 
 static int
 close_fex_sycl(VmafFeatureExtractor *fex); /* forward decl for init-failure cleanup — SY-2a */

core/src/feature/sycl/integer_vif_sycl.cpp

@@ -1373,6 +1373,7 @@ static sycl::event launch_vif_fused(sycl::queue &q, const void *ref_data, const
 static void enqueue_vif_work(void *queue_ptr, void *priv, void *shared_ref, void *shared_dis);
 static void vif_pre_graph(void *queue_ptr, void *priv);
 static void vif_post_graph(void *queue_ptr, void *priv);
+static int close_fex_sycl(VmafFeatureExtractor *fex); /* forward decl for init error paths */
 
 static int
 close_fex_sycl(VmafFeatureExtractor *fex); /* forward decl for init-failure cleanup — SY-2a */

core/test/test_gpu_dispatch_runtime.c

@@ -55,6 +55,20 @@ static int test_unsetenv(const char *name)
 #include "cuda/dispatch_strategy.h"
 #include "hip/dispatch_strategy.h"
 
+/* MinGW64 / MSVC libc lacks POSIX setenv/unsetenv; map to _putenv_s + the
+ * "VAR=" form for unset. Keeps the test source uniform across libcs. */
+#if defined(_WIN32) && !defined(__CYGWIN__)
+static int setenv(const char *name, const char *value, int overwrite)
+{
+    (void)overwrite;
+    return _putenv_s(name, value);
+}
+static int unsetenv(const char *name)
+{
+    return _putenv_s(name, "");
+}
+#endif
+
 /* ---- gpu_dispatch_parse.h — header-only static-inline parser ---- */
 
 static char *test_parse_null_inputs_decline(void)

core/test/test_y4m_alloc_failure.c

@@ -40,6 +40,7 @@
  * test_y4m_411_oob gating.
  */
 
+/* NOLINTNEXTLINE(bugprone-reserved-identifier,cert-dcl37-c,cert-dcl51-cpp) — POSIX feature-test macro */
 #define _POSIX_C_SOURCE 200809L
 
 #include <stdio.h>
@@ -56,8 +57,35 @@
  * payload is needed — the test exercises the open path only. */
 static const char kY4mAbsurdHeader[] = "YUV4MPEG2 W65535 H65535 F30:1 Ip C444p12\nFRAME\n";
 
+/* AddressSanitizer / ThreadSanitizer / MemorySanitizer reserve a multi-TB
+ * shadow virtual region on process startup. Setting RLIMIT_AS=256 MiB after
+ * that causes every subsequent ASan internal mmap to fail with "Failed to
+ * mmap" — the test process aborts before the y4m parser even runs. The
+ * regression this test guards (dst_buf-NULL bug) is C-level, not sanitizer-
+ * surfaceable, so a clean skip under sanitizers preserves coverage on the
+ * non-sanitised legs without false positives on the ASan leg. */
+#if defined(__SANITIZE_ADDRESS__) || defined(__SANITIZE_THREAD__)
+#define VMAF_TEST_SANITIZER_BUILD 1
+#elif defined(__has_feature)
+#if __has_feature(address_sanitizer) || __has_feature(thread_sanitizer) ||                         \
+    __has_feature(memory_sanitizer)
+#define VMAF_TEST_SANITIZER_BUILD 1
+#endif
+#endif
+#ifndef VMAF_TEST_SANITIZER_BUILD
+#define VMAF_TEST_SANITIZER_BUILD 0
+#endif
+
 static char *test_y4m_open_returns_error_on_oom(void)
 {
+#if VMAF_TEST_SANITIZER_BUILD
+    /* See comment above: setrlimit(RLIMIT_AS) is incompatible with sanitizer
+     * shadow reservations. The pre-fix bug this test guards is not sanitizer-
+     * detectable, so skipping under sanitizers gives up no coverage. */
+    (void)fprintf(stderr, "(sanitizer build; skipping RLIMIT_AS test) ");
+    return NULL;
+#endif
+
     /* Cap virtual address space at 256 MiB.  The header demands ≈ 25 GiB
      * which the y4m parser will try to malloc — must fail under this cap.
      * Skip the test (passing) if setrlimit isn't permitted (some sandboxed

dev/Containerfile

@@ -62,6 +62,7 @@ RUN --mount=type=cache,target=/var/cache/apt,sharing=locked \
     g++-13 \
     clang-19 \
     clang++-19 \
+    libclang-rt-19-dev \
     # Build system
     meson \
     ninja-build \

docs/rebase-notes.md

@@ -43438,3 +43438,10 @@ HIP and CUDA overviews and the metric page are fork-only docs; the
 backends index page is also fork-only. No upstream Netflix/vmaf source is
 touched. The cleanup closes residual references left over after PR #38
 (ADR-0709) removed the `float_ansnr` extractor from every backend.
+
+## fix/post-rename-post-vulkan-sweep — 2026-06-01
+
+no rebase impact: post-rename cleanup only. The Containerfile change adds a
+pkg-install line that cannot conflict with upstream (upstream has no
+Containerfile). The score_backend.py change is fork-local code with no
+upstream counterpart. The test and doc updates are purely fork-local.