chore(changelog.d): prune 7 stale fragments + rewrite 2 misleading ones

changelog.d/changed/changelog-d-stale-fragments-cleanup.md

@@ -0,0 +1,17 @@
+- **chore(changelog.d):** Prune 7 stale `changelog.d/` fragments that
+  advertised behaviour no longer on master, and rewrite 2 misleading
+  fragments. Vulkan-backend fragments (`vif-arc-mesa-anv-int64-reduction`,
+  `vulkan-vif-fp64-g-computation`, `motion-vulkan-1067-regression`) were
+  obsoleted by ADR-0726 (Vulkan backend dropped). `float_ansnr` fragments
+  (`float-ansnr-enable-chroma-pr947-restore`, `hip-ansnr-memcpy-direction`)
+  were obsoleted by the `float_ansnr` extractor removal (PR #38, cited in
+  ADR-0749). HIP `dispatch-strategy-pr1067-regression` referenced a
+  fork-history `g_hip_features[]` table that never landed on master (the
+  HIP dispatch_strategy is still the ADR-0212 stub). `metal-dispatch-strategy-key-restore`
+  was superseded by the canonical-score-names fix on master
+  (commit ff4c30d40c). `onnx-blobs-to-github-releases` was rewritten to
+  reflect the actual scaffold-only state (the fetcher script exists but
+  the 3 large blobs are still inlined in git pending the
+  `tiny-blobs-v1` Release upload). `cuda-extractor-cambi-and-ssim-promotion`
+  was rewritten to drop the contradiction with the cambi_cuda SIGSEGV
+  fixes that landed in PR #866 + PR #870.

changelog.d/changed/cuda-extractor-cambi-and-ssim-promotion.md

@@ -1,8 +1,10 @@
 - The K150K/CHUG feature extractor now routes `float_ssim` through
   the CUDA primary pass (with explicit `scale=1` per the
   `float_ssim_cuda` v1 contract) instead of the CPU residual pass.
-  `cambi` was originally also planned for promotion but stays on the
-  CPU residual: `cambi_cuda` segfaults on every input on the
-  rebuilt 2026-05-15 binary (Issue #857). Per-clip wall time on CUDA
-  workers improves by roughly the CPU SSIM cost; CAMBI remains on
-  the CPU residual until the CUDA-side bug is fixed.
+  Per-clip wall time on CUDA workers improves by roughly the CPU
+  SSIM cost. `cambi` was originally also planned for promotion;
+  the `cambi_cuda` SIGSEGV that blocked promotion (Issue #857) has
+  since been fixed (PR #866 + PR #870, 2026-05-16), but the K150K
+  script intentionally still routes `cambi` through the CPU
+  residual pass — a follow-up will re-evaluate promotion against a
+  rebuilt CUDA binary.

changelog.d/changed/onnx-blobs-to-github-releases.md

@@ -1,11 +1,10 @@
-- The three model/tiny/*.onnx blobs ≥1MB (`transnet_v2.onnx`,
-  `fastdvdnet_pre.onnx`, `lpips_sq.onnx`, totalling 44 MB) are no
-  longer inlined in git. They live as attachments on the
-  `tiny-blobs-v1` GitHub Release and are fetched on demand by
-  `scripts/ai/fetch-tiny-blobs.sh`, which sha256-verifies each
-  download against the recorded hash in
-  `model/tiny/registry.json`. New checkouts run the fetcher once
-  (~3.5 s) instead of inlining 44 MB into every clone. Per
-  ADR-0457. Smaller ONNX files (<1 MB each, 25 files) stay inline;
-  the per-file fetch overhead dominates the per-byte storage
-  savings below the cutoff.
+- Scaffolding for ADR-0457 offloading of large `model/tiny/*.onnx`
+  blobs (>=1 MB) to GitHub Release attachments: the
+  `scripts/ai/fetch-tiny-blobs.sh` fetcher script (idempotent
+  download + sha256 verification driven by `model/tiny/registry.json`)
+  is in place. The actual blob offload is not yet live — the three
+  large blobs (`transnet_v2.onnx`, `fastdvdnet_pre.onnx`,
+  `lpips_sq.onnx`, totalling ~44 MB) are still inlined in git on
+  master pending the `tiny-blobs-v1` GitHub Release upload + matching
+  registry update. Until that follow-up lands, the fetcher is a no-op
+  on fresh clones (the blobs are already present).