Skip to content

fix(ci): apply PR #767 huge-alloc deselects to tests-and-quality-gates.yml - #770

Merged
lusoris merged 1 commit into
masterfrom
fix/sanitizer-deselect-tests-and-quality-gates
Jun 6, 2026
Merged

lusoris merged 1 commit into
masterfrom
fix/sanitizer-deselect-tests-and-quality-gates

Conversation

@lusoris

@lusoris lusoris commented Jun 6, 2026

Copy link
Copy Markdown
Contributor

Summary

  • PR fix(ci): extend ASan/UBSan + TSan exclusion list for intentional huge-alloc tests #767 (bb42e36) added test_gpu_picture_pool_uaf, test_integer_motion_v2_coverage, and test_pic_preallocation to the exclusion list in sanitizers.yml only — the original ADR-0347 case-based deselect mechanism in tests-and-quality-gates.yml was left unmodified.
  • That matrix job (sanitizer: [address, undefined, thread]) was still running all three tests and SIGABRTing on the intentional ~192 GiB OOM allocation path in each.
  • This PR mirrors the same EXCLUDE additions to all three address, undefined, and thread arms of the case block in tests-and-quality-gates.yml.

Root cause

tests-and-quality-gates.yml and sanitizers.yml are two separate sanitizer gates. PR #767 fixed the new dedicated file but not the original matrix job, which uses a case "${{ matrix.sanitizer }}" shell construct (the ADR-0347 deselect mechanism) rather than inline grep -vE patterns.

Test plan

  • CI: Sanitizers — ASan + UBSan + MSan matrix job in tests-and-quality-gates.yml should no longer SIGABRT on the three OOM tests.
  • Verify: grep -vE 'test_model$|test_y4m_alloc_failure$|test_gpu_picture_pool_uaf$|test_integer_motion_v2_coverage$|test_pic_preallocation$' locally filters all 5 excluded tests and leaves the rest intact.
  • No source, header, or test file is modified — CI-only change.

Deliverables checklist

  • research digest: no digest needed: trivial one-line regex extension to CI workflow
  • decision matrix: no alternatives: only-one-way fix — add the same pattern to the case block that was already fixed in sanitizers.yml
  • AGENTS.md invariant: no rebase-sensitive invariants
  • reproducer: echo -e "libvmaf:test_gpu_picture_pool_uaf\nlibvmaf:test_psnr" | grep -vE 'test_model$|test_y4m_alloc_failure$|test_gpu_picture_pool_uaf$|test_integer_motion_v2_coverage$|test_pic_preallocation$' — passes through only libvmaf:test_psnr
  • changelog fragment: changelog.d/fixed/sanitizer-deselect-tests-quality-gates.md
  • rebase-notes: entry added

state.md

No new bug opened or closed — this is a CI-wiring fix for the gap left by PR #767; no new defect tracking row needed.

ffmpeg-patches

No public C-API or meson_options.txt change — exempt.

🤖 Generated with Claude Code

…s.yml

PR #767 (bb42e36) added test_gpu_picture_pool_uaf, test_integer_motion_v2_coverage,
and test_pic_preallocation to the exclusion list in sanitizers.yml but missed the
original ADR-0347 case-based deselect mechanism in tests-and-quality-gates.yml.
That matrix job (sanitizer: [address, undefined, thread]) was still running all
three tests and SIGABRT-ing on the intentional ~192 GiB OOM allocation path.

Add all three to the EXCLUDE pattern for the address, undefined, and thread
cases in tests-and-quality-gates.yml, mirroring exactly what sanitizers.yml now
carries. The OOM paths these tests guard remain covered by the unsanitized
meson test suite on every run.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
@lusoris
lusoris merged commit b3f53f6 into master Jun 6, 2026
43 of 57 checks passed
@lusoris
lusoris deleted the fix/sanitizer-deselect-tests-and-quality-gates branch June 6, 2026 17:50
lusoris added a commit that referenced this pull request Jun 7, 2026
Add three missing Recently-closed rows that were absent from state.md:

- T-GPU-POOL-UAF-OOM-ASAN-UBSAN-GAP-2026-06-06: PRs #767 + #770 added
  three huge-alloc tests to ASan/UBSan/TSan exclusion lists in both
  sanitizers.yml and tests-and-quality-gates.yml; CI SIGABRT spurious
  failures resolved.

- T-HIP-MOTION-DEBUG-BOOL-SYCL-GRAPH-DANGLING-2026-06-06: PR #768 fixed
  HIP motion test passing "1" for a VMAF_OPT_TYPE_BOOL option (should be
  "true") and a SYCL graph dangling-priv SIGSEGV when a VmafSyclState is
  shared across two sequential VmafContext instances.

- T-MOTION-FIVE-FRAME-WINDOW-PYTHON-SKIP-2026-06-06: PR #771 added
  @unittest.skip decorators to 9 Python test methods that set
  motion_five_frame_window=True, which returns -ENOTSUP from C per
  ADR-0337 pending prev_prev_ref plumbing.

PRs #765 (T-PREV-REF-BATCH-REFCOUNT-LEAK), #766 (T-MCP-SCORE-POOLED-EAGAIN),
and #769 (T-PIC-PREALLOC-ASAN-LEAK) were already tracked. PR #770 adds no
new bug row (CI wiring fix only, no new defect opened/closed).

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
lusoris added a commit that referenced this pull request Jun 8, 2026
…e) (#845)

* docs(internal-headers): add Doxygen @brief/@PARAM to 10 core src/*.h (ADR-1096)

214 internal headers had no Doxygen coverage; only 26 had any @brief/@PARAM
annotation. Adds @brief, @PARAM[in/out], and @return comments to the ten
highest-traffic headers:

  framesync.h thread_pool.h picture_pool.h predict.h fex_ctx_vector.h
  ref.h mem.h log.h opt.h dict.h

Purely additive — no logic, no ABI, no public-header changes.
IDE hover-docs and doxygen -q now populate for all covered APIs.

ADR-1096 documents the coverage decision and follow-up scope.
AGENTS.md invariant added: update Doxygen blocks when signatures change.

no digest needed: trivial doc-only addition
no alternatives: only-one-way fix (add the comments)

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

* docs(api): add @thread-safety/@param/@return to all undocumented public C-API functions

Every VMAF_EXPORT function in core/include/libvmaf/ now carries a complete
Doxygen contract: @PARAM, @return, and @Thread-safety. Previously libvmaf_cuda.h
(5 functions), libvmaf_sycl.h (20), dnn.h (9), picture_v2.h (5), and
model.h (vmaf_model_version_next) were missing @Thread-safety entirely;
picture_v2.h stubs were also missing @param/@return. VmafPoolingMethod in
libvmaf.h lacked a @brief enum-level description.

The thread-safety contract is uniform: GPU backend setup functions (cuda/sycl/
hip/metal state init/import/free, preallocate, fetch) are not thread-safe —
one handle per driver thread. Pure query functions (vmaf_dnn_available,
vmaf_sycl_list_devices, vmaf_dnn_verify_signature, vmaf_backend_handle_name,
vmaf_model_version_next) are marked safe from any thread.

No C source files, build files, or ABI-visible signatures changed — Doxygen
comment additions only.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

* docs(state): backfill state.md rows for PRs #765–#771

Add three missing Recently-closed rows that were absent from state.md:

- T-GPU-POOL-UAF-OOM-ASAN-UBSAN-GAP-2026-06-06: PRs #767 + #770 added
  three huge-alloc tests to ASan/UBSan/TSan exclusion lists in both
  sanitizers.yml and tests-and-quality-gates.yml; CI SIGABRT spurious
  failures resolved.

- T-HIP-MOTION-DEBUG-BOOL-SYCL-GRAPH-DANGLING-2026-06-06: PR #768 fixed
  HIP motion test passing "1" for a VMAF_OPT_TYPE_BOOL option (should be
  "true") and a SYCL graph dangling-priv SIGSEGV when a VmafSyclState is
  shared across two sequential VmafContext instances.

- T-MOTION-FIVE-FRAME-WINDOW-PYTHON-SKIP-2026-06-06: PR #771 added
  @unittest.skip decorators to 9 Python test methods that set
  motion_five_frame_window=True, which returns -ENOTSUP from C per
  ADR-0337 pending prev_prev_ref plumbing.

PRs #765 (T-PREV-REF-BATCH-REFCOUNT-LEAK), #766 (T-MCP-SCORE-POOLED-EAGAIN),
and #769 (T-PIC-PREALLOC-ASAN-LEAK) were already tracked. PR #770 adds no
new bug row (CI wiring fix only, no new defect opened/closed).

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

* chore(state): backfill 4 missing state.md rows from PRs #712–#747 batch

State.md updates promised by squash commits for PRs #723, #725, #729,
and #743 were lost (the pre-squash branch commits carried the changes
but the squash commits did not include state.md in their diffs).

Changes:
- Move T-JSON-MODEL-SLOPES-FEATURE-CAP-OOB-2026-05-30 from Open to
  Recently closed — fixed by PR #743 / ADR-0887 (vmaf_model_destroy
  heap-buffer-overflow via fuzz_json_model nightly harness).
- Add T-FFMPEG-PATCHES-SCORE-FMT-GAP-2026-06-06 to Recently closed —
  PR #723 / ADR-1064 wired score_fmt AVOption on all four FFmpeg vmaf
  filters; PR #740 fixed patch hunk counts.
- Add T-VENDORED-CJSON-PDJSON-SECURITY-2026-06-06 to Recently closed —
  PR #725 / ADR-1061 fixed five pdjson/cJSON security and correctness
  bugs (depth guard never compiled, size overflow x2, banned sprintf/
  strcpy at 12 sites, cJSON_GetArraySize int wrap).
- Add T-GO-STATICCHECK-R10-TIMER-BODY-2026-06-06 to Recently closed —
  PR #729 / ADR-1065 fixed Go timer leak (time.After in poll loop),
  missing body size cap on vmafx-controller, and missing ReadTimeout
  on both HTTP servers.
- Add second-pass _Updated: header summarising the PRs #712–#747 batch.

no rebase impact: state.md only, no code or API surface changes.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

* docs(nav): add 10 orphaned pages to mkdocs.yml nav

Audit found the following docs/**.md files present on disk but absent from
both the mkdocs.yml nav and any inbound cross-reference:

- docs/ai/sidecar-online-training.md (k8s Phase 4b sidecar trainer)
- docs/server/auth.md (JWT auth gateway for vmafx-controller)
- docs/server/operator.md (kubebuilder Kubernetes operator)
- docs/server/rest.md (vmafx-server REST/OpenAPI surface)
- docs/development/ebpf-fuse-bypass.md (rclone FUSE eBPF bypass)
- docs/development/perf-claims-2026-05-10.md (May 2026 perf claims log)
- docs/sync-upstream/2026-05-02-sync-report.md (upstream sync report)
- docs/sync-upstream/2026-05-03-sync-report.md (upstream sync report)
- docs/upstream-ports/1b08bb4d-needs-manual-port.md (manual port note)

Also excludes docs/changelog.d/** from the mkdocs build — one changelog
fragment (cpp23-wave3.md) was placed under docs/changelog.d/ instead of
the repo-root changelog.d/; the fragment is referenced by the ADR index
but is not a standalone page and should not be rendered by mkdocs.

No code changes. No ADR required (nav-only housekeeping).

no rebase impact: docs/mkdocs.yml nav entries only
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

* fix(helm): values completeness — nameOverride, statePVCSize, node.metricsPort, extraPorts schema (ADR-1074)

Four gaps where values.yaml / values.schema.json diverged from template usage:

1. nameOverride/fullnameOverride: read by _helpers.tpl but absent from both
   values.yaml and the root additionalProperties:false schema; any user
   supplying --set nameOverride=foo received an immediate helm-lint failure.
   Added as string keys to both files.

2. statefulSet.statePVCSize: statefulset.yaml hardcoded `storage: 1Gi` for the
   per-replica MCP-state PVC. Exposed as statefulSet.statePVCSize (default 1Gi)
   and wired into the volumeClaimTemplates storage field.

3. node.metricsPort: port 9090 appeared hardcoded in three template locations
   (node Deployment containerPort, node-metrics Service port, NetworkPolicy
   allow rule). Exposed as node.metricsPort (default 9090) and unified.

4. service.extraPorts items schema: bare `"type": "array"` with no items
   definition accepted malformed port objects silently. Added items schema with
   required [name, port] and protocol enum.

All defaults preserve existing rendered output byte-for-byte.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

* test(ai/scripts): add unit tests for calibrate_phase_f_recipes and analyze_knob_sweep helpers (round 3)

The existing test_calibrate_phase_f_recipes.py covers only the main()
invocation (run-provenance smoke).  The existing test_knob_sweep_analysis.py
covers pareto_frontier, stratify, and detect_recipe_regressions via a
20-row synthetic fixture.

This commit adds coverage for the remaining pure helper functions:

* test_calibrate_phase_f_recipes_unit.py (50 tests):
  - mos_to_vmaf_proxy: clamping, boundary MOS values, string coercion
  - saliency_benefit_to_intensity: threshold boundaries for all three labels
  - _iter_corpus_rows: valid rows, blank-line skip, malformed-JSON skip,
    missing-field skip, optional duration_s default
  - _ugc_target_vmaf_offset: empty/small corpus, symmetric distribution,
    heavy-tail sign, clamp bounds, rounding
  - _ugc_tight_interval_width: empty corpus fallback, uniform/wide
    distributions, floor/cap enforcement, rounding
  - _resolution_dominance: empty, single-resolution, split, dominant bucket,
    portrait-vs-landscape distinction
  - _ugc_saliency_benefit_fraction: fallback, no-qualify, all-qualify,
    half-qualify, high-MOS exclusion, square-aspect treatment
  - calibrate() integration: all four recipe classes, UGC/proxy provenance
    tags, required recipe keys, saliency label validity

* test_analyze_knob_sweep_unit.py (29 tests):
  - _stable_knob_repr: empty dict, single entry, alphabetical sort, non-Mapping
    input, numeric values, insertion-order invariance
  - _slug: alphanumeric passthrough, hyphen/underscore preserved, space/slash
    replacement, empty-string fallback, all-special-chars
  - _closest_bare_at_bitrate: no-bare-rows, within tolerance, outside
    tolerance, picks closest, exact match, zero-tolerance
  - write_slice_csv: filename pattern, header row, data rows, slug-safe names,
    auto-creates output directory
  - write_summary_md: file creation, slice count, no-regression message,
    regression table, auto-creates output directory

All 79 tests run without GPU, corpus, or model downloads (<100 ms total).

no rebase impact: test-only addition; no existing golden assertion modified.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

* test(vmafx-server): replace //nolint:errcheck with explicit discards per ADR-0278

Replace all //nolint:errcheck suppression comments in grpc_server_handler_test.go
with explicit-discard patterns (_ = x.Close() / defer func() { _ = x.Close() }()).
Also remove duplicate TestGRPCScore_ScorerError function (the earlier copy was
accidentally left in; keep the one with the fuller doc comment at the bottom of
the file). Remove duplicate TestRunHTTP_BadAddress from main_extra_test.go.
Fix unchecked ln.Close() and conn.Close() in TestRunHTTPGracefulShutdown.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

---------

Co-authored-by: Lusoris <lusoris@pm.me>
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
@lusoris lusoris added this to the 1.0.0 — First release milestone Sep 4, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant