lusoris · May 31, 2026 · May 31, 2026
@@ -4,10 +4,10 @@ The fork's `CHANGELOG.md` Unreleased block is **rendered**, not edited
 directly. Each PR adds one fragment file under one of the
 Keep-a-Changelog sections:
 
-```
+```text
 changelog.d/
   added/        new user-discoverable surface
-  changed/      enhancement to an existing surface
+  changed/      enhancement to an existing surface (includes perf — name files perf-<topic>.md)
   deprecated/   surface marked for removal
   removed/      surface deleted
   fixed/        bug fix with user-visible delta
@@ -17,6 +17,13 @@ changelog.d/
                             section at the next release tag)
 ```
 
+There is **no `perf/` or `performance/` directory** — the renderer
+warns on any non-Keep-a-Changelog subdirectory and skips its
+contents. Performance fragments live in `changed/` with a `perf-`
+filename prefix so they sort contiguously inside the rendered
+`### Changed` section. See [ADR-0892](../docs/adr/0892-conventional-commits-and-changelog-fragment-hygiene.md)
+and [ADR-0913](../docs/adr/0913-changelog-renderer-splice-contract.md).
+
 ## How to add a fragment
 
 1. Pick the section directory matching your change (Keep-a-Changelog).

@@ -1,5 +1,3 @@
-## Added
-
 - **`speed_qa` real SpEED-QA implementation** (`core/src/feature/speed_qa.c`):
   replaces the 0.0-placeholder scaffold with a working spatial and temporal
   entropic-differencing extractor per Bampis et al. 2017 (IEEE SPL 24(9)).

@@ -1,5 +1,3 @@
-## Added
-
 - **Tiny-AI Netflix corpus training scaffold — draft PR registration (ADR-0417).**
   Opens `ai/tiny-netflix-training-scaffold` as a formal draft PR so the user can
   review and confirm architecture choices before triggering a training run against

@@ -1,4 +1,4 @@
-## `integer_vif`: add `enable_chroma` option
+### `integer_vif`: add `enable_chroma` option
 
 `integer_vif` now accepts `enable_chroma` (bool, default `false`). When enabled,
 the four-scale VIF pipeline runs on the Cb and Cr planes in addition to luma and

@@ -1,4 +1,4 @@
-## vmaf-tune executor: per-shot and saliency execution modes (ADR-0468)
+### vmaf-tune executor: per-shot and saliency execution modes (ADR-0468)
 
 `vmaftune.executor` gains two new execution-mode entry-points that extend the Phase F
 base execute mode (ADR-0454):

@@ -19,7 +19,7 @@ clip; before this PR the HIP integer_motion path never executed any
 HIP code because `compute_fex_flags()` didn't include the HIP slot
 and the per-extractor flag bit was cleared.
 
-## What ships
+### What ships
 
 - New picture-buffer enum entry `VMAF_PICTURE_BUFFER_TYPE_HIP_DEVICE`
   reserved for the future HIP picture pool. Today's HIP TUs accept
@@ -48,7 +48,7 @@ and the per-extractor flag bit was cleared.
   `hip_kernel_sources` map (previously the source file existed but
   was not compiled or linked).
 
-## What does NOT ship
+### What does NOT ship
 
 `vmaf_fex_integer_vif_hip` was speculatively flagged in its batch-1
 commit but crashes with a GPU memory access fault on the first
@@ -59,7 +59,7 @@ float_motion / float_ssim / float_psnr / cambi / float_adm /
 ssimulacra2) stay unflagged pending per-extractor end-to-end
 verification.
 
-## ADRs
+### ADRs
 
 - [ADR-0530](../docs/adr/0530-hip-feature-flag-promotion-and-picture-buffer.md)
   — this change.

@@ -1,5 +1,3 @@
-## Added
-
 - **Real `two_pass_args` for all 14 codec adapters that previously raised
   `NotImplementedError`** (ADR-0546). Software adapters `libaom-av1`
   and `libvvenc` now run true two-invocation 2-pass via FFmpeg's

@@ -1,5 +1,3 @@
-## Added
-
 - **Cross-backend parity matrix** (`docs/research/0550-cross-backend-parity-matrix-2026-05-18.md`):
   systematic audit of all 18 CPU feature extractors across SYCL (Intel Arc A380),
   CUDA (RTX 4090), and Vulkan on the Netflix 576x324 golden fixture.

@@ -1,5 +1,3 @@
-## Added
-
 - **`--tiny-resize {bilinear,nearest,bicubic,disabled}` CLI flag** plus
   matching public C API `vmaf_dnn_set_resize_mode(ctx, mode)` and
   `VmafDnnResizeMode` enum
@@ -18,7 +16,7 @@
   replicate-edge clamping. Bit-identical to `vmaf_tensor_from_luma`
   on the no-resize fast path. ADR-0550.
 
-## Fixed
+### Fixed
 
 - **NR tiny-model 0-frames footgun (post-fix probe Finding 11)**:
   `vmaf --no-reference --tiny-model nr_metric_v1.onnx --distorted

@@ -1,4 +1,4 @@
-## MCP P1 surface: vmaf-tune integration, list_extractors, describe_model, progress notifications (ADR-0608)
+### MCP P1 surface: vmaf-tune integration, list_extractors, describe_model, progress notifications (ADR-0608)
 
 Five new MCP tools added to `vmaf-mcp`:
 

@@ -1,5 +1,3 @@
-## Added
-
 - **`.zed/` project configuration** — Zed editor parity with `.vscode/` (ADR-0608):
   `.zed/settings.json` (clangd LSP, pyright, ruff, shfmt, vmaf-mcp MCP server,
   Claude Agent via ACP, file-type associations for CUDA/HIP/Metal/GLSL, telemetry

@@ -1,4 +1,4 @@
-## `vmaf-tune --fast-nr`: NR early-elimination for 2–4× bisect wall-time cut
+### `vmaf-tune --fast-nr`: NR early-elimination for 2–4× bisect wall-time cut
 
 Added `--fast-nr` flag to `vmaf-tune compare` and `vmaf-tune tune-per-shot`
 (ADR-0624, implementing ADR-0615).  At each bisect midpoint the cheap

@@ -1,5 +1,3 @@
-## Added
-
 - **`ai/scripts/validate_chug_hdr_mos_head.py`**: held-out test-partition
   validator for the CHUG HDR MOS head. Loads a CHUG MOS head ONNX, filters
   CHUG feature JSONL shards to `split == "test"` rows (552 rows, never used

@@ -1,5 +1,3 @@
-## Added
-
 - **ADR-0447** — documents the decision to correct motion-feature under-reporting
   on HFR (50p+) content by applying `motion_fps_weight = clamp(30/fps, 0.25, 4.0)`
   across all motion extractor variants (CPU, CUDA, SYCL, Vulkan). Closes Issue #837

@@ -1,5 +1,3 @@
-## Added
-
 - **`cambi` in cross-backend parity gate** (`scripts/ci/cross_backend_parity_gate.py`):
   adds `cambi` to the `FEATURE_METRICS` and `FEATURE_TOLERANCE` dicts with the metric
   name `Cambi_feature_cambi_score` and places=4 tolerance (5e-5). Closes coverage gap

@@ -1,4 +1,4 @@
-## Added: Runtime resolution-aware CUDA kernel variant dispatch (ADR-0753)
+### Added: Runtime resolution-aware CUDA kernel variant dispatch (ADR-0753)
 
 A new `vmaf_cuda_workload_class(w, h)` classifier (`core/src/feature/cuda/resolution_dispatch.{h,c}`)
 maps each frame's luma pixel count to `WS_SMALL` (< 720p), `WS_MEDIUM` (720p–4K), or

@@ -1,5 +1,3 @@
-## Added
-
 - **`scripts/dev/cleanup-agent-state.sh`**: developer utility that removes
   stale agent worktrees (locked with a dead PID) and drops redundant git
   stashes (on `master`, detached HEAD, or branches that still exist locally).

@@ -4,7 +4,7 @@ Promotes two HIP feature extractors from `-ENOSYS` scaffolds to real
 HIP Module API consumers, following the pattern established by PR #612
 (`float_psnr_hip`, ADR-0254).
 
-## What ships
+### What ships
 
 - **`psnr_score.hip`** — device kernel for `integer_psnr_hip`:
   uint64 atomic-SSE reduction, warp-size-64 `__shfl_down` (split into
@@ -23,7 +23,7 @@ HIP Module API consumers, following the pattern established by PR #612
   (`hipcc --genco` + `xxd -i`), mirroring the CUDA `ptx_arrays` pattern.
   Controlled by the existing `enable_hipcc` option.
 
-## How to use
+### How to use
 
 ```bash
 # Build with real HIP kernels (requires ROCm 6+ and hipcc in PATH):
@@ -39,7 +39,7 @@ vmaf --reference ref.yuv --distorted dis.yuv --feature float_ansnr_hip \
      --width 1920 --height 1080 --pixel_format 420
 ```
 
-## Skipped (batch-2 candidates)
+### Skipped (batch-2 candidates)
 
 - `float_ssim_hip`: two-pass design with five intermediate device
   buffers and three kernel functions — non-trivial ABI adaptation.

@@ -1,4 +1,4 @@
-## HIP float_adm real kernel (ADR-0468)
+### HIP float_adm real kernel (ADR-0468)
 
 Port the CUDA `float_adm_cuda.c` / `float_adm_score.cu` four-stage DWT+CSF+CM pipeline
 to HIP as the ninth kernel-template consumer. Adds `float_adm_hip.c`,

@@ -1,5 +1,3 @@
-## Added
-
 - **`float_ansnr_hip` `enable_chroma` option** (ADR-0453 parity):
   `FloatAnsnrStateHip` gains `bool enable_chroma` (default `false`) and
   `unsigned n_planes`. When `enable_chroma=true`, the extractor dispatches

@@ -1,5 +1,3 @@
-### Added
-
 - HIP (AMD ROCm) backend runtime — flips
   `core/src/hip/{common,kernel_template}.c` from the audit-first
   `-ENOSYS` scaffold (ADR-0212) to real HIP runtime calls. New

@@ -4,7 +4,7 @@ Promotes the seven remaining Metal feature-extractor scaffolds from
 `-ENOSYS` stubs to functional MSL kernels, following the `integer_motion_v2`
 template established in T8-1c.
 
-## What ships
+### What ships
 
 - **`float_psnr.metal`** / **`float_psnr_metal.mm`** (T8-1d): per-pixel
   `(ref - dis)^2` float accumulation. 8bpc and 16bpc kernel variants.
@@ -44,14 +44,14 @@ template established in T8-1c.
   intermediate blurred planes in MTLBuffers allocated at init. Features:
   `float_ssim`, `float_ms_ssim`.
 
-## Key design note: no `atomic_ulong`
+### Key design note: no `atomic_ulong`
 
 All kernels use per-WG `float`/`uint` partials arrays — **not** MSL
 `atomic_fetch_add_explicit` for 64-bit types, which fails silently on
 Apple Silicon (CI run 25685703780 / job 75408804495). Host reduces in
 `double`. See `docs/research/0421-metal-kernel-batch-t8-1c-k.md`.
 
-## Build (macOS only)
+### Build (macOS only)
 
 ```bash
 meson setup build libvmaf -Denable_metal=enabled \
@@ -60,7 +60,7 @@ ninja -C build
 meson test -C build test_metal_smoke
 ```
 
-## Parity gate
+### Parity gate
 
 ```bash
 scripts/ci/cross_backend_parity_gate.py \

@@ -1,5 +1,3 @@
-## Added
-
 - **`scripts/git-hooks/pre-push-mkdocs-strict.sh` — mkdocs strict-mode pre-push gate
   ([ADR-0466](../docs/adr/0466-mkdocs-strict-pre-push-hook.md)).**
   Detects whether a push touches `docs/` or `mkdocs.yml`; skips immediately on

@@ -1,5 +1,3 @@
-## Added
-
 - **Model cards for 4 previously undocumented tiny-AI checkpoints**
   (`docs/ai/models/vmaf_tiny_v1.md`, `vmaf_tiny_v1_medium.md`,
   `smoke_v0.md`, `smoke_fp16_v0.md`): closes the ADR-0042 F5 gap

@@ -1,5 +1,3 @@
-### Added
-
 - **`float_moment` SVE2 SIMD path** (`core/src/feature/arm64/moment_sve2.c`):
   the 1st- and 2nd-moment reductions now use the VLA (vector-length-agnostic)
   SVE2 f32→f64 widening pattern on Linux aarch64 hosts that expose

@@ -1,4 +1,4 @@
-## vmafx-node: rclone remote-asset streaming (Phase 4b.5)
+### vmafx-node: rclone remote-asset streaming (Phase 4b.5)
 
 `vmafx-node` now bundles rclone and streams reference and distorted video
 directly from remote storage (S3, GCS, Azure Blob, SFTP, HTTP, and 70+ other

@@ -1,4 +1,4 @@
-## Added multi-resolution performance benchmark baseline
+### Added multi-resolution performance benchmark baseline
 
 `scripts/perf/bench-multi-resolution.sh` is now the canonical multi-resolution,
 multi-backend, multi-metric throughput benchmark for VMAF.  It runs across

@@ -1,5 +1,3 @@
-## Added
-
 - **SpEED Python compat wrappers** (`compat/python-vmaf/core/feature_extractor.py`):
   ports `SpeedChromaFeatureExtractor` and `SpeedTemporalFeatureExtractor` from
   Netflix/vmaf upstream into the compat harness. Both wrap the `speed_chroma` and

@@ -1,5 +1,3 @@
-## Added
-
 - `float_ssim_sycl`: add `enable_chroma` option (default `false`) and `n_planes`
   field to `SsimStateSycl`, mirroring the CUDA twin (PR #950). The v1 kernel
   is luma-only; `n_planes` is clamped to 1. Multi-plane dispatch is deferred

@@ -1,5 +1,3 @@
-### Added
-
 - **`--tiny-codec` / `--tiny-preset` / `--tiny-crf` CLI flags** populate
   the codec one-hot block of codec-aware tiny models (today
   `fr_regressor_v2`) so the model receives the real encoder context

@@ -1,4 +1,4 @@
-## vmaf-tune: `--neg` flag for VMAF NEG model variants (ADR-0622)
+### vmaf-tune: `--neg` flag for VMAF NEG model variants (ADR-0622)
 
 Add a `--neg` flag to `vmaf-tune recommend`, `compare`, `tune-per-shot`,
 `ladder`, and `corpus`. When set, the flag routes the VMAF model to the

@@ -1,4 +1,4 @@
-## `vmaf-tune auto` — three additional F.1/F.2 short-circuit predicates
+### `vmaf-tune auto` — three additional F.1/F.2 short-circuit predicates
 
 Three new short-circuit predicates (#8, #9, #10) are appended to
 `SHORT_CIRCUIT_PREDICATES` in `tools/vmaf-tune/src/vmaftune/auto.py`:

@@ -1,5 +1,3 @@
-## Added
-
 - **`core/test/dnn/test_vmaf_use_tiny_model.c`**: 5 new unit tests covering
   the `vmaf_use_tiny_model()` public ctx-attach entry point, which had zero
   C-unit-test coverage (identified in `audit-test-coverage-2026-05-16.md §2`).

@@ -1,4 +1,4 @@
-## `vmafx` binary and AI tool aliases (ADR-0690)
+### `vmafx` binary and AI tool aliases (ADR-0690)
 
 - Install `vmafx` as a symlink to `vmaf` in the same `bindir` via Meson
   `install_symlink()`. No separate binary; one binary, one symlink.

@@ -1,4 +1,4 @@
-## vmafx-mcp: Go implementation of the VMAFX MCP server
+### vmafx-mcp: Go implementation of the VMAFX MCP server
 
 Added `cmd/vmafx-mcp/` — a single static Go binary that exposes the same 15
 MCP tools as the Python `vmaf-mcp` server with byte-for-byte schema parity

@@ -1,4 +1,4 @@
-## `--netflix-compat` flag for legacy-default opt-out (ADR-0696)
+### `--netflix-compat` flag for legacy-default opt-out (ADR-0696)
 
 - Add `--netflix-compat` / `--netflix_compat` CLI flag to both the `vmaf` and
   `vmafx` binaries.

@@ -1,4 +1,4 @@
-## VMAFX Phase 4 language-modernization foundation (ADR-0702)
+### VMAFX Phase 4 language-modernization foundation (ADR-0702)
 
 - Added Go workspace root (`go.mod`, module `github.com/VMAFx/vmafx`, Go 1.23)
   with `pkg/version` package and `go-ci.yml` CI gate.

@@ -1,4 +1,4 @@
-## vmafx-sys Rust FFI crate
+### vmafx-sys Rust FFI crate
 
 Add `bindings/rust/vmafx-sys` — bindgen-generated raw FFI bindings to libvmaf plus
 a thin safe Rust wrapper layer (`vmafx_sys::safe`). The safe layer exposes

@@ -1,4 +1,4 @@
-## vmafx-tune-go Stage 1 — Go port of vmaf-tune compare subcommand
+### vmafx-tune-go Stage 1 — Go port of vmaf-tune compare subcommand
 
 Adds `cmd/vmafx-tune/` — a Go port of the Python `vmaf-tune` rate-quality
 tuning CLI, installed as `vmafx-tune-go` alongside the Python binary.

@@ -1,4 +1,4 @@
-## vmafx-tune-go Stage 2 — `ladder` subcommand (per-title ABR bitrate-ladder)
+### vmafx-tune-go Stage 2 — `ladder` subcommand (per-title ABR bitrate-ladder)
 
 Adds `vmafx-tune-go ladder` — the per-title ABR bitrate-ladder generator,
 porting the Python `vmaf-tune ladder` Phase E algorithm to Go.

@@ -1,4 +1,4 @@
-## Vulkan backend: persistent `VkPipelineCache` (ADR-0445)
+### Vulkan backend: persistent `VkPipelineCache` (ADR-0445)
 
 The Vulkan compute backend now persists its compiled pipeline cache to
 `$XDG_CACHE_HOME/libvmaf/vulkan-pipeline-cache.bin` across process

@@ -1,4 +1,4 @@
-## chore(feature): delete orphan HIP/CUDA TUs (ADR-0546)
+### chore(feature): delete orphan HIP/CUDA TUs (ADR-0546)
 
 Removed 6 dead/orphan translation units found by a deep audit of
 `core/src/feature/hip/` and `core/src/feature/cuda/`:

@@ -1,4 +1,4 @@
-## Deduplicate file-hash and UTC-timestamp helpers in vmaf-tune corpus
+### Deduplicate file-hash and UTC-timestamp helpers in vmaf-tune corpus
 
 `tools/vmaf-tune/src/vmaftune/corpus.py` previously contained private inline
 implementations of `_sha256_of` (streaming SHA-256 file hash) and `_utc_now_iso`

@@ -1,5 +1,3 @@
-## Changed
-
 - **CUDA toolkit pin bumped to 13.3** (`dev/Containerfile`):
   `cuda-toolkit-13-2` → `cuda-toolkit-13-3` (13.3.0-1 from NVIDIA ubuntu2404
   apt repo, GA 2026-05-28). Minimum host driver updated to R610.43.02 (Linux

@@ -1,4 +1,4 @@
-## chore(ci): add concurrency groups + shell-strict to curl|tar steps
+### chore(ci): add concurrency groups + shell-strict to curl|tar steps
 
 Adds top-level `concurrency:` blocks with `cancel-in-progress: true` to three
 CI workflows that lacked them, so stale runs are cancelled when a new push or

@@ -1,4 +1,4 @@
-## CI matrix slimmed to 1 build per OS + state-of-the-art sanitizers (ADR-0710)
+### CI matrix slimmed to 1 build per OS + state-of-the-art sanitizers (ADR-0710)
 
 `libvmaf-build-matrix.yml` (15 build rows post-ADR-0689) is replaced by
 `build.yml` with three matrix rows:

@@ -1,5 +1,3 @@
-### Changed
-
 - **CLAUDE.md §12 r12**: dropped the stale "T7-5 NOLINT sweep is pending"
   paragraph since PR #327 (refactor pass) and PR #388 (citation closeout,
   ADR-0278) discharged that backlog item; every NOLINT in tree now carries

@@ -1,5 +1,3 @@
-### Changed
-
 - CodeQL C bulk sweep: fixed 60 of 78 deferred alerts across 18 upstream-mirror
   files. Categories addressed: `cpp/integer-multiplication-cast-to-long` (44 of
   44 alerts; widening cast prefixed on the LHS operand before multiplication),

@@ -1,4 +1,4 @@
-## chore(review): adversarial code review — C++23 wave (PRs #41–#58)
+### chore(review): adversarial code review — C++23 wave (PRs #41–#58)
 
 Read-only adversarial review of the cpp23 conversion wave. Found 4 CRITICAL, 2 HIGH,
 10 MEDIUM, 3 LOW issues across all 9 PRs. Key findings: