lusoris · Oct 5, 2026 · Oct 5, 2026
@@ -4288,6 +4288,17 @@ make `core/AGENTS.md` a generated index over `AGENTS.d/` topic pages ([ADR-1454]
   combination stops the operator at startup.
 
 
+- **The vendored Pelorus x265 CSV reader builds without warnings under an
+  optimising GCC.** `x265_csv_read_rows()` in
+  `core/src/interop/pelorus_qp_report_csv.c` left its column-index struct
+  uninitialised until the header row was seen, and GCC 16 at `-O2 -Wall -Wextra`
+  raised seven `-Wmaybe-uninitialized` warnings (`type`, `poc`, `qp`, `bits`,
+  `psnr_y`, `psnr_u`, `psnr_v`). Every index now starts at -1 (absent), which the
+  reader already treats as a missing column. The fix is in VMAFx/pelorus
+  (#79) and re-vendored here (pin `42cb17106a2d`); the parsed values are
+  unchanged.
+
+
 - **The ASan + UBSan job no longer kills `test_pic_preallocation`.** The test
   runs the `vmaf_v0.6.1` model on 1080p frames in the unoptimised sanitizer
   build, about 9 s of CPU, and a hosted runner running the fast suite in

@@ -0,0 +1,9 @@
+- **The vendored Pelorus x265 CSV reader builds without warnings under an
+  optimising GCC.** `x265_csv_read_rows()` in
+  `core/src/interop/pelorus_qp_report_csv.c` left its column-index struct
+  uninitialised until the header row was seen, and GCC 16 at `-O2 -Wall -Wextra`
+  raised seven `-Wmaybe-uninitialized` warnings (`type`, `poc`, `qp`, `bits`,
+  `psnr_y`, `psnr_u`, `psnr_v`). Every index now starts at -1 (absent), which the
+  reader already treats as a missing column. The fix is in VMAFx/pelorus
+  (#79) and re-vendored here (pin `42cb17106a2d`); the parsed values are
+  unchanged.
@@ -17,7 +17,7 @@
  */
 
 /*
- * VENDORED FROM VMAFx/pelorus@013bc59f04f1701d9f9e8e6ff2a2f6b985c6d955 — DO NOT EDIT.
+ * VENDORED FROM VMAFx/pelorus@42cb17106a2d3fae7790754f7cd8c6e1fbe6fa7f — DO NOT EDIT.
  * Append-only ABI; single
  * source of truth is pelorus. Re-sync via scripts/sync-pelorus-interop.sh.
  * See docs/adr/1113-vendor-pelorus-interop-abi.md.

@@ -17,7 +17,7 @@
  */
 
 /*
- * VENDORED FROM VMAFx/pelorus@013bc59f04f1701d9f9e8e6ff2a2f6b985c6d955 — DO NOT EDIT.
+ * VENDORED FROM VMAFx/pelorus@42cb17106a2d3fae7790754f7cd8c6e1fbe6fa7f — DO NOT EDIT.
  * Append-only ABI; single
  * source of truth is pelorus. Re-sync via scripts/sync-pelorus-interop.sh.
  * See docs/adr/1113-vendor-pelorus-interop-abi.md.

@@ -17,7 +17,7 @@
  */
 
 /*
- * VENDORED FROM VMAFx/pelorus@013bc59f04f1701d9f9e8e6ff2a2f6b985c6d955 — DO NOT EDIT.
+ * VENDORED FROM VMAFx/pelorus@42cb17106a2d3fae7790754f7cd8c6e1fbe6fa7f — DO NOT EDIT.
  * Append-only ABI; single
  * source of truth is pelorus. Re-sync via scripts/sync-pelorus-interop.sh.
  * See docs/adr/1113-vendor-pelorus-interop-abi.md.

@@ -17,7 +17,7 @@
  */
 
 /*
- * VENDORED FROM VMAFx/pelorus@013bc59f04f1701d9f9e8e6ff2a2f6b985c6d955 — DO NOT EDIT.
+ * VENDORED FROM VMAFx/pelorus@42cb17106a2d3fae7790754f7cd8c6e1fbe6fa7f — DO NOT EDIT.
  * Append-only ABI; single
  * source of truth is pelorus. Re-sync via scripts/sync-pelorus-interop.sh.
  * See docs/adr/1113-vendor-pelorus-interop-abi.md.

@@ -17,7 +17,7 @@
  */
 
 /*
- * VENDORED FROM VMAFx/pelorus@013bc59f04f1701d9f9e8e6ff2a2f6b985c6d955 — DO NOT EDIT.
+ * VENDORED FROM VMAFx/pelorus@42cb17106a2d3fae7790754f7cd8c6e1fbe6fa7f — DO NOT EDIT.
  * Append-only ABI; single
  * source of truth is pelorus. Re-sync via scripts/sync-pelorus-interop.sh.
  * See docs/adr/1113-vendor-pelorus-interop-abi.md.

@@ -17,7 +17,7 @@
  */
 
 /*
- * VENDORED FROM VMAFx/pelorus@013bc59f04f1701d9f9e8e6ff2a2f6b985c6d955 — DO NOT EDIT.
+ * VENDORED FROM VMAFx/pelorus@42cb17106a2d3fae7790754f7cd8c6e1fbe6fa7f — DO NOT EDIT.
  * Append-only ABI; single
  * source of truth is pelorus. Re-sync via scripts/sync-pelorus-interop.sh.
  * See docs/adr/1113-vendor-pelorus-interop-abi.md.

@@ -17,7 +17,7 @@
  */
 
 /*
- * VENDORED FROM VMAFx/pelorus@013bc59f04f1701d9f9e8e6ff2a2f6b985c6d955 — DO NOT EDIT.
+ * VENDORED FROM VMAFx/pelorus@42cb17106a2d3fae7790754f7cd8c6e1fbe6fa7f — DO NOT EDIT.
  * Append-only ABI; single
  * source of truth is pelorus. Re-sync via scripts/sync-pelorus-interop.sh.
  * See docs/adr/1113-vendor-pelorus-interop-abi.md.

@@ -17,7 +17,7 @@
  */
 
 /*
- * VENDORED FROM VMAFx/pelorus@013bc59f04f1701d9f9e8e6ff2a2f6b985c6d955 — DO NOT EDIT.
+ * VENDORED FROM VMAFx/pelorus@42cb17106a2d3fae7790754f7cd8c6e1fbe6fa7f — DO NOT EDIT.
  * Append-only ABI; single
  * source of truth is pelorus. Re-sync via scripts/sync-pelorus-interop.sh.
  * See docs/adr/1113-vendor-pelorus-interop-abi.md.
@@ -425,7 +425,9 @@ static pel_result x265_csv_read_rows(FILE *fp, PelorusX265Frame *out_frames, siz
 {
     char line[PEL_CSV_LINE_MAX]; /* bounded, fixed (Po10): no heap, no VLA  */
     char *fields[PEL_CSV_MAX_FIELDS];
-    csv_cols cols;
+    /* Every index starts "absent" (-1): a row is then never admitted before a header has set the
+     * real indices, and the compiler sees a defined value on every path. */
+    csv_cols cols = {-1, -1, -1, -1, -1, -1, -1};
 
     while (fgets(line, (int)sizeof(line), fp) != NULL) {
         size_t nf = split_fields(line, fields, PEL_CSV_MAX_FIELDS);

@@ -17,7 +17,7 @@
  */
 
 /*
- * VENDORED FROM VMAFx/pelorus@013bc59f04f1701d9f9e8e6ff2a2f6b985c6d955 — DO NOT EDIT.
+ * VENDORED FROM VMAFx/pelorus@42cb17106a2d3fae7790754f7cd8c6e1fbe6fa7f — DO NOT EDIT.
  * Append-only ABI; single
  * source of truth is pelorus. Re-sync via scripts/sync-pelorus-interop.sh.
  * See docs/adr/1113-vendor-pelorus-interop-abi.md.

@@ -2743,7 +2743,7 @@ libvmaf_sources = [
     # ADR-1113: vendored Pelorus <-> vmafx interop ABI (pack/parse, deband +
     # denoise param contracts, x265 CSV QP-report reader, version/result-string
     # accessors). CPU-only, dependency-free, NO Vulkan. Verbatim mirror of
-    # VMAFx/pelorus@013bc59f04f1701d9f9e8e6ff2a2f6b985c6d955 (ABI 1.3).
+    # VMAFx/pelorus@42cb17106a2d3fae7790754f7cd8c6e1fbe6fa7f (ABI 1.3).
     # DO NOT EDIT; re-sync via
     # scripts/sync-pelorus-interop.sh. See docs/api/pelorus-interop.md.
     src_dir + 'interop/pelorus_interop.c',

@@ -18,7 +18,7 @@
 
 /*
  * test_pelorus_interop.c — vmafx side of the SHARED Pelorus interop ABI
- * conformance fixture (VMAFx/pelorus@013bc59f04f1701d9f9e8e6ff2a2f6b985c6d955
+ * conformance fixture (VMAFx/pelorus@42cb17106a2d3fae7790754f7cd8c6e1fbe6fa7f
  * test/interop_test.c, ABI 1.3).
  *
  * Both repos run byte-for-byte the same checks against their own copy of

@@ -18,7 +18,7 @@ re-pin and exact-mirror guard rails are recorded in
 
 !!! warning "The mirror is read-only"
     Do not edit the vendored files. They are byte-identical to their Pelorus
-    origin (pinned at `VMAFx/pelorus@013bc59f04f1701d9f9e8e6ff2a2f6b985c6d955`)
+    origin (pinned at `VMAFx/pelorus@42cb17106a2d3fae7790754f7cd8c6e1fbe6fa7f`)
     except for a `VENDORED FROM ... DO NOT EDIT` banner and the include-path
     rewrite described below. Fix any defect upstream in Pelorus, then
     [re-sync](#re-syncing-the-mirror).
@@ -50,7 +50,7 @@ Pelorus's. (`qp_report_csv.c` is required to link: the ABI-1.3 fixture exercises
 ### The only local edits
 
 1. A `VENDORED FROM
-   VMAFx/pelorus@013bc59f04f1701d9f9e8e6ff2a2f6b985c6d955 — DO NOT EDIT`
+   VMAFx/pelorus@42cb17106a2d3fae7790754f7cd8c6e1fbe6fa7f — DO NOT EDIT`
    banner inserted after the (unchanged) Pelorus license header.
 2. Intra-Pelorus `#include "pelorus/<x>.h"` rewritten to
    `#include "libvmaf/pelorus/<x>.h"` so the headers resolve under
@@ -200,7 +200,7 @@ python3 scripts/ci/run_meson_test.py -- \
 The pin and the drift guard live in
 [`scripts/sync-pelorus-interop.sh`](../../scripts/sync-pelorus-interop.sh). It
 reads the vendored sources from the **pinned commit's git tree object**
-(`git show 013bc59f04f1701d9f9e8e6ff2a2f6b985c6d955:libpelorus/…`), so it
+(`git show 42cb17106a2d3fae7790754f7cd8c6e1fbe6fa7f:libpelorus/…`), so it
 stays accurate even when the local Pelorus checkout's `HEAD` has moved past the
 pin. A directory that is not a Git checkout, or a checkout that lacks the exact
 object, fails closed.
@@ -215,7 +215,7 @@ scripts/sync-pelorus-interop.sh /path/to/pelorus
 ```
 
 The output is
-`OK   - mirror matches pelorus@013bc59f04f1701d9f9e8e6ff2a2f6b985c6d955`,
+`OK   - mirror matches pelorus@42cb17106a2d3fae7790754f7cd8c6e1fbe6fa7f`,
 or `FAIL - mirror has drifted` with a diff and exit code 1.
 
 To re-vendor after a reviewed Pelorus ABI addition or a released parser

@@ -61652,3 +61652,11 @@ No score, public API or FFmpeg patch impact.
   pin (ADR-1113); `compat/python-vmaf/core/adm_dwt2_cy.pyx` keeps its first line,
   `# SPDX-License-Identifier: BSD-2-Clause-Patent`, when upstream Netflix is synced. No score,
   public API or FFmpeg patch impact.
+
+## Pelorus pin moves to the qp_report_csv initialisation fix
+
+- `PELORUS_VENDOR_SHA` moves to `42cb17106a2d` (VMAFx/pelorus #79: `csv_cols` starts at -1 in
+  `x265_csv_read_rows()`). The mirror stays verbatim (ADR-1113): a conflict in
+  `core/src/interop/pelorus_*.c` takes master's side and re-runs
+  `scripts/sync-pelorus-interop.sh --update`; never merge a hunk by hand. No score, public API
+  or FFmpeg patch impact.
@@ -1022,6 +1022,7 @@ landed fix yet._
 ## Recently closed
 
 | **T-GOLDEN-ASSERTIONS-BEHIND-UPSTREAM-2026-10-05** — the fork's golden assertions held Netflix's older values at looser places than Netflix's own re-records (`5c7770080`, `005988ead`, `4679db83c`, `d93495f5c`, `e3827e4dd`: 164 changed assertions), because the golden-data rule read as a ban on any edit | **FOUND by the upstream sync of 2026-10-05 and FIXED on `port/upstream-golden-updates-2026-05` (opened and closed by one PR).** A measurement against the fork's CPU build found 156 of 164 reproduced at upstream's places, none failing, 8 not exercised. The maintainer decided to adopt them; the rule now allows exactly this, a verbatim port of Netflix's own update after a measurement. `make test-netflix-golden` passes on the PR head. | [ADR-1828](adr/1828-port-netflix-golden-updates.md) | `port/upstream-golden-updates-2026-05` | 2026-10-05 | closed |
+| **T-PELORUS-QP-CSV-MAYBE-UNINIT-2026-10-05** — the vendored `core/src/interop/pelorus_qp_report_csv.c` gave seven `-Wmaybe-uninitialized` warnings at gcc 16 `-O2 -Wall -Wextra` (`cols.type`, `poc`, `qp`, `bits`, `psnr_y`, `psnr_u`, `psnr_v`, lines 217-290): `x265_csv_read_rows()` left its `csv_cols` uninitialised until the header row; HISS-10 binds vendored code | **FOUND by the RC3 hygiene read of the #2117 re-vendor and FIXED on `fix/pelorus-qp-csv-maybe-uninit` (opened and closed by one PR).** The cause is fixed in VMAFx/pelorus ([#79](https://github.com/VMAFx/pelorus/pull/79), merge `42cb17106a2d`: every index starts at -1, no pragma; its CI gains an optimised gcc werror build) and re-vendored here with `scripts/sync-pelorus-interop.sh --update`. Verified: the file at gcc 16.2.1 `-std=c11 -O2 -Wall -Wextra` gave 7 warnings on the old pin and gives 0 on the new one; `--check` reports no drift. Known remainder: no vmafx CI lane builds with `-Werror` (`core/meson.build` sets `warning_level=2` only), so a new warning in any file is not caught here. | none: bug fix | `fix/pelorus-qp-csv-maybe-uninit` | `scripts/sync-pelorus-interop.sh <pelorus checkout>; gcc -std=c11 -O2 -Wall -Wextra -Icore/include -Icore/src -c core/src/interop/pelorus_qp_report_csv.c -o /dev/null` |
 | **T-TESTER-REPORT-DROPS-FAILURE-CAUSE-2026-10-05** — the tester report threw away the evidence that names a failure: a failed `vmaf` run kept only its last stderr line, and a unit-test program that died on a signal was a bare `fail` | **FOUND in the Apple M4 Pro report of [#2118](https://github.com/VMAFx/vmafx/issues/2118) and FIXED on `fix/tester-keep-failure-diagnostics` (opened and closed by one PR).** In that report the Metal equivalence errors on both 1080p checkerboards read `vmaf exited 234: libvmaf WARNING est_params: covariance matrix was singular on 4 of 4 solves`: a warning printed while closing, after the message that named the `-EINVAL` (`tools/rc1-tester/src/vmaf_rc1_tester/hw_equiv.py` kept `stderr.splitlines()[-1:]` only). The same report listed `test_metal_ssimulacra2_parity` as `fail` with eight passing cases and no message: the program died on a signal in its ninth case (`T-METAL-SSIMULACRA2-YUV400-ACCEPTED-2026-10-05`), and `hw_suites.py` recorded neither the signal nor the case; a timeout dropped the output it had already printed. Now a failed run's `error` keeps, after the head line (`vmaf exited N`, the signal's name for a crash, the last line), every distinct `problem ...`, `error: ...` and libvmaf `ERROR` / `WARNING` line (at most 20 lines, 4 KB); a unit-test program killed by a signal, timed out, stopped at the output limit, or exiting with a failure status without a failing case gets a line in `unit_tests.reason` (`<test>: killed by signal 11 (SIGSEGV) during case <case>`), and in a program that prints `@case` lines the case that started without a verdict is `fail` with `no verdict printed: ...`; a timeout keeps the cases printed before it. The report schema is unchanged (schema 3's `error`, `reason`, `cases`, `case_messages`). `tests/test_hw_report.py` and `tests/test_hw_metal.py` hold ten new cases, and all ten fail on master's `hw_equiv.py` / `hw_suites.py`. | [docs/usage/tester-image.md](usage/tester-image.md), [ADR-1496](adr/1496-metal-gate-in-tester-bundle.md) | `fix/tester-keep-failure-diagnostics` | `python3 -m pytest -q tools/rc1-tester/tests/test_hw_report.py tools/rc1-tester/tests/test_hw_metal.py` |
 | **T-SPEED-FLOAT-GATE-DEFAULT-MODEL-2026-10-05** — a `-Denable_float=false` build could not score with the default model: `speed_chroma` and `speed_temporal` were compiled and registered only with `enable_float=true`, and `vmaf_v1.0.16_3d0h` reads `speed_chroma` (`could not initialize feature extractor "Speed_chroma_feature_speed_chroma_uv_score"`, `problem loading feature extractors from model: vmaf_v1.0.16_3d0h`) | **FOUND by the upstream sync of 2026-10-05 (Netflix/vmaf `6046b1926` and the build hunk of `4718b4f5f` were not ported) and FIXED on `port/6046b1926-speed-without-float` (opened and closed by one PR).** `speed.c`, `speed_internal.c`, `vif_tools.c` and `common/convolution.c` move to the unconditional source list of `core/src/meson.build`, the two extractors out of `#if VMAF_FLOAT_FEATURES` in `core/src/feature/feature_extractor.cpp`, and the SpEED tests lose their `enable_float` gate. Measured on `ryzen-4090-arc`, GCC 16.2.1, release builds without LTO: on master `52e265fc0` with `-Denable_float=false` the default model fails on the Netflix 576x324 pair; after the change it scores (pooled `vmaf` mean 82.81606015944988) and its JSON equals the `-Denable_float=true` build's byte for byte apart from `fps`/`version`. Float build before and after at `--precision max`, the three Netflix pairs: default model and `speed_chroma` + `speed_temporal` + `float_vif` + `float_motion` reports identical (6 of 6, 108 frames, 1404 values). `test_speed` fails in a float-off build when the registry hunk is reverted (`speed_chroma extractor must be registered by name`) and passes with it; the seven SpEED test binaries pass in the float-off build; fast suite 327 of 327; Netflix golden gate 280 passed, 3 skipped. | (none: upstream port) | `port/6046b1926-speed-without-float` | 2026-10-05 | closed |
 | **T-HARDWARE-NEEDS-CPU-ROW-OVERALL-VERDICT-2026-10-05** — the "Hardware we need" table (`docs/usage/hardware-we-need.md`, `scripts/docs/generate-hardware-reports.py`) gave a CPU family row the report's overall verdict, so a report whose CPU checks all passed and whose GPU section failed rated the processor row "worst fail" | **FOUND on the office workstation's regeneration of the table and on the outside-tester reports of 2026-10-05 (#2116: UHD 770 SYCL image, every CPU check passing, `failed_checks` `["gpu"]`), and FIXED on `fix/hardware-needs-cpu-row-verdict` (opened and closed by one PR).** `_verdicts()` appended `report["verdict"]` for every host-matched row. `_host_verdict()` now rates a CPU row from the report's CPU checks (dispatch and reference equivalence, unit tests, golden check) and `image.files_match_build`, through `CHECK_KEYS` / `PASSING` of `vmaf_rc1_tester.hw_report`; the native macOS row adds the Metal equivalence and the Metal gate (it closes the Metal rows). Run over the three reports of 2026-10-05 (#2116, #2118, #2119), the AVX2 row reads "2 reported, worst pass" (was "worst fail"); the Apple, Ampere and Xe-LP rows are unchanged (fail, pass, fail). Test: `CpuRowVerdictTests` in `scripts/docs/tests/test_hardware_needs.py` (`test_gpu_failure_does_not_fail_the_cpu_row` fails on the old generator). | — | `fix/hardware-needs-cpu-row-verdict` | 2026-10-05 | fixed |

@@ -44,7 +44,7 @@ set -euo pipefail
 # reviewed interop ABI addition or a parser correctness/security fix, even when
 # ABI major/minor stay unchanged. Keep this in lock step with every vendored
 # banner and docs/api/pelorus-interop.md.
-PELORUS_VENDOR_SHA="013bc59f04f1701d9f9e8e6ff2a2f6b985c6d955"
+PELORUS_VENDOR_SHA="42cb17106a2d3fae7790754f7cd8c6e1fbe6fa7f"
 
 repo_root="$(git rev-parse --show-toplevel 2>/dev/null || pwd)"
 mirror_policy="$repo_root/scripts/ci/pelorus_mirror.py"