fix: resolve all remaining 24bb5daf89 conflict markers (70 files, supersedes #181)

ai/src/aiutils/jsonl_utils.py

@@ -5,7 +5,6 @@
 from __future__ import annotations
 
 import json
-<<<<<<< HEAD
 import math
 from pathlib import Path
 from typing import Any, Iterator
@@ -44,10 +43,6 @@ def dumps_jsonl_row(obj: dict, **kwargs: Any) -> str:
     sanitised = _sanitize_nonfinite(obj)
     kwargs.setdefault("sort_keys", True)
     return json.dumps(sanitised, **kwargs) + "\n"
-=======
-from pathlib import Path
-from typing import Iterator
->>>>>>> 24bb5daf89 (docs: post-merge-train sweep — VMAFx + core/ path refs, ADR index, state.md)
 
 
 def iter_jsonl(path: Path) -> Iterator[tuple[int, dict]]:

ai/src/vmaf_train/registry.py

@@ -17,10 +17,7 @@
 
 import hashlib
 import json
-<<<<<<< HEAD
 import math
-=======
->>>>>>> 24bb5daf89 (docs: post-merge-train sweep — VMAFx + core/ path refs, ADR index, state.md)
 from dataclasses import asdict, dataclass, field
 from pathlib import Path
 from typing import Any
@@ -117,7 +114,6 @@ def register(
 def load(sidecar_path: Path) -> ModelMetadata:
     doc: dict[str, Any] = json.loads(sidecar_path.read_text())
     return ModelMetadata(**doc)
-<<<<<<< HEAD
 
 
 def _sanitize_nonfinite(obj: Any) -> Any:
@@ -168,5 +164,3 @@ def write_registry_json(path: Path, payload: dict, **kwargs: Any) -> None:
         **kwargs: Forwarded to :func:`dumps_registry_json` / :func:`json.dumps`.
     """
     path.write_text(dumps_registry_json(payload, **kwargs) + "\n", encoding="utf-8")
-=======
->>>>>>> 24bb5daf89 (docs: post-merge-train sweep — VMAFx + core/ path refs, ADR index, state.md)

core/AGENTS.md

@@ -457,7 +457,6 @@ the corrected methodology.
   — `enable_avx512=true` with `enable_asm=false` issues a warning (no-op, not an error);
   — `enable_hipcc=true` with `enable_hip=false` issues a warning (no-op, not an error).
   The checks run at configuration time (before `subdir()` calls) to catch misconfigurations early. The principle: every option that depends on another must `error()` on the bad combo, never silently no-op. See [`src/meson.build` lines 100–111, 74–76, 142–144](src/meson.build).
-<<<<<<< HEAD
 
 - **C→C++23 conversion safety invariants** (adversarial review 2026-05-28,
   `docs/research/cpp23-wave-adversarial-review-20260528.md`):
@@ -513,8 +512,6 @@ the corrected methodology.
   The legacy path (`VmafFeatureExtractor`, line ~301) retains the mapping as
   documented debt — tracked as T-LEGACY-RUNNER-ANSNR-BROKEN in `docs/state.md`.
   The checks run at configuration time (before `subdir()` calls) to catch misconfigurations early. The principle: every option that depends on another must `error()` on the bad combo, never silently no-op. See [`src/meson.build` lines 100–111, 74–76, 142–144](src/meson.build).
-=======
->>>>>>> 24bb5daf89 (docs: post-merge-train sweep — VMAFx + core/ path refs, ADR index, state.md)
 
 ## Performance benchmark invariant (ADR-0752)
 

docs/adr/_index_fragments/0425-vmaf-roi-score-saliency-materialiser.md

@@ -1,5 +1 @@
-<<<<<<< HEAD
 | [ADR-0425](../0425-vmaf-roi-score-saliency-materialiser.md) | vmaf-roi-score saliency materialiser | Accepted | 2026-05-14 |
-=======
-| [ADR-0425](0425-vmaf-roi-score-saliency-materialiser.md) | vmaf-roi-score saliency materialiser | Accepted | 2026-05-14 |
->>>>>>> 24bb5daf89 (docs: post-merge-train sweep — VMAFx + core/ path refs, ADR index, state.md)

docs/adr/_index_fragments/0578-vif-scratch-buf-hoist-to-vifstate.md

@@ -1,5 +1 @@
-<<<<<<< HEAD
 | [ADR-0578](0452-vif-scratch-buf-hoist-to-vifstate.md) | Hoist VIF 10-plane scratch buffer from per-frame allocation to VifState init/close lifecycle; eliminates ~79 MB/frame allocator traffic at 1080p | Accepted | perf, vif, cpu, build, fork-local |
-=======
-| [ADR-0578](0578-vif-scratch-buf-hoist-to-vifstate.md) | Hoist VIF 10-plane scratch buffer from per-frame allocation to VifState init/close lifecycle; eliminates ~79 MB/frame allocator traffic at 1080p | Accepted | perf, vif, cpu, build, fork-local |
->>>>>>> 24bb5daf89 (docs: post-merge-train sweep — VMAFx + core/ path refs, ADR index, state.md)

docs/adr/_index_fragments/0583-float-ms-ssim-enable-chroma.md

@@ -1,5 +1 @@
-<<<<<<< HEAD
 | [ADR-0583](0461-float-ms-ssim-enable-chroma.md) | Add `enable_chroma` option to the `float_ms_ssim` extractor | Accepted | ms-ssim, float-ms-ssim, option-parity, metrics, correctness, fork-local |
-=======
-| [ADR-0583](0583-float-ms-ssim-enable-chroma.md) | Add `enable_chroma` option to the `float_ms_ssim` extractor | Accepted | ms-ssim, float-ms-ssim, option-parity, metrics, correctness, fork-local |
->>>>>>> 24bb5daf89 (docs: post-merge-train sweep — VMAFx + core/ path refs, ADR index, state.md)

docs/adr/_index_fragments/0589-metal-ssim-lcs-db-parity.md

@@ -1,5 +1 @@
-<<<<<<< HEAD
 | [ADR-0589](0484-metal-ssim-lcs-db-parity.md) | Metal `float_ssim` option parity — `enable_lcs`, `enable_db`, `clip_db`, `scale` (full CPU parity) | Accepted | metal, ssim, option-parity, apple-silicon, kernel, fork-local |
-=======
-| [ADR-0589](0589-metal-ssim-lcs-db-parity.md) | Metal `float_ssim` option parity — `enable_lcs`, `enable_db`, `clip_db`, `scale` (full CPU parity) | Accepted | metal, ssim, option-parity, apple-silicon, kernel, fork-local |
->>>>>>> 24bb5daf89 (docs: post-merge-train sweep — VMAFx + core/ path refs, ADR index, state.md)

docs/adr/_index_fragments/0594-hip-ssimulacra2-blur-fp-contract-off.md

@@ -1,5 +1 @@
-<<<<<<< HEAD
 | [ADR-0594](0539-hip-ssimulacra2-blur-fp-contract-off.md) | Per-kernel `hip_cu_extra_flags` dict in `core/src/meson.build` (first entry `ssimulacra2_blur: ['-ffp-contract=off']`) mirrors the CUDA `cuda_cu_extra_flags --fmad=false` carve-out so the recursive Gaussian IIR `n2*sum - d1*prev` step keeps IEEE-754 add/mul ordering on hipcc / amdclang++ device code and `ssimulacra2` HIP scores stay inside the ADR-0214 places=2 parity gate vs CPU. Fall-through (`.get(name, [])`) is byte-identical for every kernel not listed. Verified on AMD gfx1036 — `ssimulacra2` / `integer_ssim` / `integer_ms_ssim` HIP scores match CPU to display precision. | Accepted | hip, build, ssimulacra2, numerics, fork-local |
-=======
-| [ADR-0594](0594-hip-ssimulacra2-blur-fp-contract-off.md) | Per-kernel `hip_cu_extra_flags` dict in `core/src/meson.build` (first entry `ssimulacra2_blur: ['-ffp-contract=off']`) mirrors the CUDA `cuda_cu_extra_flags --fmad=false` carve-out so the recursive Gaussian IIR `n2*sum - d1*prev` step keeps IEEE-754 add/mul ordering on hipcc / amdclang++ device code and `ssimulacra2` HIP scores stay inside the ADR-0214 places=2 parity gate vs CPU. Fall-through (`.get(name, [])`) is byte-identical for every kernel not listed. Verified on AMD gfx1036 — `ssimulacra2` / `integer_ssim` / `integer_ms_ssim` HIP scores match CPU to display precision. | Accepted | hip, build, ssimulacra2, numerics, fork-local |
->>>>>>> 24bb5daf89 (docs: post-merge-train sweep — VMAFx + core/ path refs, ADR index, state.md)

docs/adr/_index_fragments/0600-upstream-port-direct-read.md

@@ -1,5 +1 @@
-<<<<<<< HEAD
 | [ADR-0600](0567-upstream-port-direct-read.md) | Port Netflix upstream `30a6e2a8d` — add `USE_DIRECT_READ` compile flag to CLI tools that eliminates the intermediate `video_input_ycbcr` buffer and per-frame `memcpy` by reading directly into preallocated `VmafPicture` planes via a new `fetch_into_vmaf_picture` vtable slot. Extends `vidinput.h/c`, `yuv_input.c`, `y4m_input.c`, and `vmaf.c`; Y4M direct path gated on `convert == y4m_convert_null`. Flag is opt-in; old path retained under `#else`. | Accepted | upstream-port, performance, tools, cli, build |
-=======
-| [ADR-0600](0600-upstream-port-direct-read.md) | Port Netflix upstream `30a6e2a8d` — add `USE_DIRECT_READ` compile flag to CLI tools that eliminates the intermediate `video_input_ycbcr` buffer and per-frame `memcpy` by reading directly into preallocated `VmafPicture` planes via a new `fetch_into_vmaf_picture` vtable slot. Extends `vidinput.h/c`, `yuv_input.c`, `y4m_input.c`, and `vmaf.c`; Y4M direct path gated on `convert == y4m_convert_null`. Flag is opt-in; old path retained under `#else`. | Accepted | upstream-port, performance, tools, cli, build |
->>>>>>> 24bb5daf89 (docs: post-merge-train sweep — VMAFx + core/ path refs, ADR index, state.md)

docs/ai/model-registry.md

@@ -106,11 +106,7 @@ deployments should set it on. `cosign` must be on `$PATH`; install
 prebuilt binaries from the [Sigstore release page](https://github.com/sigstore/cosign/releases).
 
 The C entry point is `vmaf_dnn_verify_signature(onnx_path, registry_path)`
-<<<<<<< HEAD
 in [`core/include/libvmaf/dnn.h`](../../core/include/libvmaf/dnn.h);
-=======
-in [`core/include/core/dnn.h`](../../core/include/core/dnn.h);
->>>>>>> 24bb5daf89 (docs: post-merge-train sweep — VMAFx + core/ path refs, ADR index, state.md)
 both arguments are NULL-tolerant in the documented way.
 
 ## Directory jail — `VMAF_TINY_MODEL_DIR`

docs/ai/multi-corpus-aggregation.md

@@ -34,11 +34,7 @@ distribution shape is preserved.
 | `lsvq`             | 1.0–5.0 ACR Likert                 | `unified = (mos - 1) * 25`  | Ying et al., CVPR 2021 §4.1 — https://github.com/baidut/PatchVQ |
 | `youtube-ugc`      | 1.0–5.0 ACR Likert                 | `unified = (mos - 1) * 25`  | Wang et al., MMSP 2019 §3.2 — https://media.withyoutube.com/ |
 | `waterloo-ivc-4k`  | 0–100 continuous (DCR-like)        | identity                    | Cheon & Lee, CVPR-W 2016 §III.B — https://ece.uwaterloo.ca/~zduanmu/cvpr2016_4kvqa/ |
-<<<<<<< HEAD
 | `netflix-public`   | VMAF 0–100 (objective proxy)       | identity                    | `core/include/libvmaf/model.h` |
-=======
-| `netflix-public`   | VMAF 0–100 (objective proxy)       | identity                    | `core/include/core/model.h` |
->>>>>>> 24bb5daf89 (docs: post-merge-train sweep — VMAFx + core/ path refs, ADR index, state.md)
 
 The mapping is a single source of truth in
 `ai/scripts/aggregate_corpora.py:SCALE_CONVERSIONS`; the unit tests

docs/ai/overview.md

@@ -4,11 +4,7 @@ The **Tiny AI** surface lets you ship small, specialized perceptual-quality
 models alongside the classic VMAF SVM without introducing a second ML
 runtime or giving up libvmaf's C-only deployment story. The whole feature
 is gated on `-Denable_dnn=auto|enabled|disabled` (default `auto`) and
-<<<<<<< HEAD
 consumed through a single public header, `libvmaf/dnn.h`.
-=======
-consumed through a single public header, `core/dnn.h`.
->>>>>>> 24bb5daf89 (docs: post-merge-train sweep — VMAFx + core/ path refs, ADR index, state.md)
 
 ## The four capabilities
 

docs/ai/security.md

@@ -132,11 +132,7 @@ the `--tiny-model-verify` flag invokes `cosign verify-blob` at load time
 via `posix_spawnp(3p)` and fails closed if the signature is missing or
 bad. Off by default for dev-friendliness; strongly recommended on for
 production deployments. The flag drives `vmaf_dnn_verify_signature()` in
-<<<<<<< HEAD
 [`core/include/libvmaf/dnn.h`](../../core/include/libvmaf/dnn.h),
-=======
-[`core/include/core/dnn.h`](../../core/include/core/dnn.h),
->>>>>>> 24bb5daf89 (docs: post-merge-train sweep — VMAFx + core/ path refs, ADR index, state.md)
 which looks up the model's `sigstore_bundle` field in
 [`model/tiny/registry.json`](../../model/tiny/registry.json) — see
 [model-registry.md](model-registry.md) for the full schema and CLI flow.

docs/api/dnn.md

@@ -1,12 +1,6 @@
-<<<<<<< HEAD
 # DNN session API — `libvmaf/dnn.h`
 
 The DNN surface in [`core/include/libvmaf/dnn.h`](../../core/include/libvmaf/dnn.h)
-=======
-# DNN session API — `core/dnn.h`
-
-The DNN surface in [`core/include/core/dnn.h`](../../core/include/core/dnn.h)
->>>>>>> 24bb5daf89 (docs: post-merge-train sweep — VMAFx + core/ path refs, ADR index, state.md)
 lets callers load and run tiny ONNX models from C, either attached to a
 `VmafContext` (so DNN scores show up next to SVM scores in the normal VMAF
 report) or as a standalone session (luma-in / luma-out filter-style
@@ -313,11 +307,7 @@ Only works when libvmaf was built with `-Denable_dnn=true`.
 
 The fork ships a Sigstore-keyless verification primitive for tiny-AI
 ONNX bundles, surfaced by both the C API
-<<<<<<< HEAD
 (`vmaf_dnn_verify_signature` in `core/include/libvmaf/dnn.h`) and
-=======
-(`vmaf_dnn_verify_signature` in `core/include/core/dnn.h`) and
->>>>>>> 24bb5daf89 (docs: post-merge-train sweep — VMAFx + core/ path refs, ADR index, state.md)
 the CLI flag `--tiny-model-verify` on the `vmaf` binary
 (see [`docs/usage/cli.md`](../usage/cli.md)).
 

docs/api/gpu.md

@@ -35,11 +35,7 @@ Backend dispatch rules + runtime precedence:
 
 ### Header
 
-<<<<<<< HEAD
 [`core/include/libvmaf/libvmaf_cuda.h`](../../core/include/libvmaf/libvmaf_cuda.h)
-=======
-[`core/include/core/libvmaf_cuda.h`](../../core/include/core/libvmaf_cuda.h)
->>>>>>> 24bb5daf89 (docs: post-merge-train sweep — VMAFx + core/ path refs, ADR index, state.md)
 
 ### Lifecycle addition
 
@@ -213,11 +209,7 @@ int main(void) {
 
 ### Header
 
-<<<<<<< HEAD
 [`core/include/libvmaf/libvmaf_sycl.h`](../../core/include/libvmaf/libvmaf_sycl.h)
-=======
-[`core/include/core/libvmaf_sycl.h`](../../core/include/core/libvmaf_sycl.h)
->>>>>>> 24bb5daf89 (docs: post-merge-train sweep — VMAFx + core/ path refs, ADR index, state.md)
 
 ### State
 
@@ -425,11 +417,7 @@ the enable/disable pair to gate which frame ranges get timed.
 
 ### Header
 
-<<<<<<< HEAD
 [`core/include/libvmaf/libvmaf_vulkan.h`](../../core/include/libvmaf/libvmaf_vulkan.h)
-=======
-[`core/include/core/libvmaf_vulkan.h`](../../core/include/core/libvmaf_vulkan.h)
->>>>>>> 24bb5daf89 (docs: post-merge-train sweep — VMAFx + core/ path refs, ADR index, state.md)
 
 ### State
 

docs/api/index.md

@@ -1,29 +1,20 @@
 # Public C API reference
 
-<<<<<<< HEAD
 libvmaf ships a stable C API under [`core/include/libvmaf/`](../../core/include/libvmaf/).
-=======
-libvmaf ships a stable C API under [`core/include/core/`](../../core/include/core/).
->>>>>>> 24bb5daf89 (docs: post-merge-train sweep — VMAFx + core/ path refs, ADR index, state.md)
 This page is the canonical reference for the *core* API (context / picture /
 feature / model). GPU-backend entry points and the DNN session API each get
 their own page:
 
 - [core](index.md) — this page
 - [gpu.md](gpu.md) — `libvmaf_cuda.h`, `libvmaf_sycl.h`, `libvmaf_vulkan.h`,
   `libvmaf_hip.h`, `libvmaf_metal.h`
-<<<<<<< HEAD
 - [dnn.md](dnn.md) — `libvmaf/dnn.h` (tiny-AI ONNX session)
-=======
-- [dnn.md](dnn.md) — `core/dnn.h` (tiny-AI ONNX session)
->>>>>>> 24bb5daf89 (docs: post-merge-train sweep — VMAFx + core/ path refs, ADR index, state.md)
 - [mcp.md](mcp.md) — `libvmaf_mcp.h` (embedded MCP server)
 
 ## What each header exposes
 
 | Header | Symbols | Purpose |
 | --- | --- | --- |
-<<<<<<< HEAD
 | [`libvmaf.h`](../../core/include/libvmaf/libvmaf.h) | `VmafContext`, `VmafConfiguration`, lifecycle + scoring functions | Main entry point. Everything else is pulled in transitively. |
 | [`picture.h`](../../core/include/libvmaf/picture.h) | `VmafPicture`, `VmafPixelFormat`, alloc / unref | Per-frame pixel container (YUV planes + metadata). |
 | [`feature.h`](../../core/include/libvmaf/feature.h) | `VmafFeatureDictionary` | Key/value options passed to a feature extractor. |
@@ -37,21 +28,6 @@ their own page:
 | [`libvmaf_mcp.h`](../../core/include/libvmaf/libvmaf_mcp.h) | `VmafMcpServer`, `VmafMcpConfig`, transport start/stop | Embedded MCP server. Only usable in a build with `-Denable_mcp=true`. [Deep dive](mcp.md). |
 | [`vmaf_assert.h`](../../core/include/libvmaf/vmaf_assert.h) | `VMAF_ASSERT*` macros | Internal assertion helpers. Not for public use — may disappear. |
 | [`version.h`](../../core/include/libvmaf/libvmaf.h) (generated) | `VMAF_VERSION_MAJOR` etc. | Compile-time version constants. Run-time: `vmaf_version()`. |
-=======
-| [`libvmaf.h`](../../core/include/core/libvmaf.h) | `VmafContext`, `VmafConfiguration`, lifecycle + scoring functions | Main entry point. Everything else is pulled in transitively. |
-| [`picture.h`](../../core/include/core/picture.h) | `VmafPicture`, `VmafPixelFormat`, alloc / unref | Per-frame pixel container (YUV planes + metadata). |
-| [`feature.h`](../../core/include/core/feature.h) | `VmafFeatureDictionary` | Key/value options passed to a feature extractor. |
-| [`model.h`](../../core/include/core/model.h) | `VmafModel`, `VmafModelConfig`, `VmafModelCollection*` | Classic SVM model + bootstrap model collection. |
-| [`dnn.h`](../../core/include/core/dnn.h) | `VmafDnnSession`, `VmafDnnConfig`, tiny-model attach | Tiny-AI (ONNX Runtime) surface. [Deep dive](dnn.md). |
-| [`libvmaf_cuda.h`](../../core/include/core/libvmaf_cuda.h) | `VmafCudaState`, CUDA picture prealloc | CUDA backend. Only usable in a build with `-Denable_cuda=true`. [Deep dive](gpu.md#cuda). |
-| [`libvmaf_sycl.h`](../../core/include/core/libvmaf_sycl.h) | `VmafSyclState`, zero-copy frame buffers, dmabuf / VA / D3D11 import | SYCL backend. Only usable in a build with `-Denable_sycl=true`. [Deep dive](gpu.md#sycl). |
-| [`libvmaf_vulkan.h`](../../core/include/core/libvmaf_vulkan.h) | `VmafVulkanState`, queue / device lifecycle, zero-copy `VkImage` import | Vulkan compute backend. Only usable in a build with `-Denable_vulkan=true`. [Deep dive](gpu.md#vulkan). |
-| [`libvmaf_hip.h`](../../core/include/core/libvmaf_hip.h) | `VmafHipState`, lifecycle, picture prealloc | AMD HIP/ROCm backend. Only usable in a build with `-Denable_hip=true`. [Deep dive](gpu.md#hip). |
-| [`libvmaf_metal.h`](../../core/include/core/libvmaf_metal.h) | `VmafMetalState`, lifecycle, IOSurface import | Apple Metal backend. Runtime, IOSurface import, and the first eight feature kernels are usable in a build with `-Denable_metal=auto/enabled` on Apple Silicon; unsupported hosts return `-ENODEV`. [Deep dive](gpu.md#metal). |
-| [`libvmaf_mcp.h`](../../core/include/core/libvmaf_mcp.h) | `VmafMcpServer`, `VmafMcpConfig`, transport start/stop | Embedded MCP server. Only usable in a build with `-Denable_mcp=true`. [Deep dive](mcp.md). |
-| [`vmaf_assert.h`](../../core/include/core/vmaf_assert.h) | `VMAF_ASSERT*` macros | Internal assertion helpers. Not for public use — may disappear. |
-| [`version.h`](../../core/include/core/libvmaf.h) (generated) | `VMAF_VERSION_MAJOR` etc. | Compile-time version constants. Run-time: `vmaf_version()`. |
->>>>>>> 24bb5daf89 (docs: post-merge-train sweep — VMAFx + core/ path refs, ADR index, state.md)
 
 All declarations are C (with `extern "C"` guards for C++ callers). The fork has
 no C++ entry points in its public API.

docs/api/mcp.md

@@ -4,11 +4,7 @@ The fork ships an in-process MCP (Model Context Protocol) server that an
 embedding host (e.g. an editor plugin or a measurement-orchestration
 daemon) can drive over loopback HTTP, a Unix-domain socket, or a
 caller-owned stdio fd pair. The C surface is defined in
-<<<<<<< HEAD
 [`core/include/libvmaf/libvmaf_mcp.h`](../../core/include/libvmaf/libvmaf_mcp.h).
-=======
-[`core/include/core/libvmaf_mcp.h`](../../core/include/core/libvmaf_mcp.h).
->>>>>>> 24bb5daf89 (docs: post-merge-train sweep — VMAFx + core/ path refs, ADR index, state.md)
 
 This page covers the **embedded C API**. The standalone Python MCP
 server under [`mcp-server/vmaf-mcp/`](../../mcp-server/vmaf-mcp/) is a

docs/api/vulkan-image-import.md

@@ -12,11 +12,7 @@ and [ADR-0186](../adr/0186-vulkan-image-import-impl.md) (impl +
 ffmpeg-patch wiring).
 
 The public header is
-<<<<<<< HEAD
 [`core/include/libvmaf/libvmaf_vulkan.h`](../../core/include/libvmaf/libvmaf_vulkan.h).
-=======
-[`core/include/core/libvmaf_vulkan.h`](../../core/include/core/libvmaf_vulkan.h).
->>>>>>> 24bb5daf89 (docs: post-merge-train sweep — VMAFx + core/ path refs, ADR index, state.md)
 The Vulkan backend overview lives in
 [`docs/backends/vulkan/overview.md`](../backends/vulkan/overview.md).
 
@@ -130,11 +126,7 @@ on the host.
 
 - [`docs/backends/vulkan/overview.md`](../backends/vulkan/overview.md)
   — backend overview, kernel matrix, build flags.
-<<<<<<< HEAD
 - [`core/include/libvmaf/libvmaf_vulkan.h`](../../core/include/libvmaf/libvmaf_vulkan.h)
-=======
-- [`core/include/core/libvmaf_vulkan.h`](../../core/include/core/libvmaf_vulkan.h)
->>>>>>> 24bb5daf89 (docs: post-merge-train sweep — VMAFx + core/ path refs, ADR index, state.md)
   — authoritative API surface.
 - [`docs/usage/ffmpeg.md`](../usage/ffmpeg.md) — `vf_libvmaf_vulkan`
   filter that consumes the import API end-to-end.