Research-0090: Upstream SKIP-bucket triage — 2026-05-08¶

• Date: 2026-05-08
• Author: Lusoris
• Status: Companion to Research-0089 (parent survey)
• Tags: upstream-sync, port-only, triage, motion_v2, cambi, python-test, fork-local

Context¶

Research-0089 ran the /sync-upstream skill against the most recent 50 upstream commits (since 2026-04-08). The Pass-1 subject match plus Pass-2 added-identifier heuristic split the candidates into:

• 9 commits with newly-introduced top-level identifiers (function defs, struct types, macros) — those landed in the parent survey's UNPORTED / PARTIAL / PORTED-SILENTLY buckets and have already been classified.
• 41 commits with no new top-level identifiers — silently binned as SKIP-doc-or-format by the heuristic.

The latter bucket is the subject of this digest. The heuristic lives in .claude/skills/sync-upstream/SKILL.md and is deliberately conservative; "no new identifier" is not the same as "no semantic change". Several commits in this bucket modify hot-path bodies (cambi histogram layout, motion_v2 mirror boundary), add public-API fields (prev_prev_ref), or mass-port new test scaffolding (MyTestCase migration, ~3000 LOC of fixtures).

This digest classifies each of the 41 SHAs into one of five buckets so the maintainer can act on the substantive subset without re-reading 41 diffs.

Method¶

For each SHA in the 41-commit SKIP set:

1. Read git show <sha> (subject + diff).
2. For every touched path, probe fork master with git ls-tree origin/master <path> to confirm the file exists, then diff selected hunks vs the fork tree to detect silent ports.
3. For substantive C-code changes, search the fork for identifiers / patterns introduced by the commit (git grep -E … origin/master). A high hit ratio with a matching commit subject in git log origin/master indicates a silent port.
4. Cross-reference open in-flight worktree branches (feat/upstream-port-motion-v2-cluster-2026-05-08, feat/upstream-port-cambi-cluster-2026-05-08, feat/port-upstream-py-test-mytestcase-2026-05-08, docs/upstream-port-cambi-high-res-speedup-2026-05-08) to decide whether a commit is already in-progress.

The classification taxonomy (taken from the task brief):

• MERGE_BOUNDARY — Merge pull request #N from …. Captured by the underlying squashed commits; not portable.
• PORT_NOW — substantive change that should be ported.
• PORT_LATER — substantive but blocked on something.
• DEFER_INDEFINITELY — touches Netflix-golden tests in a way the fork doesn't want, or modifies stable surfaces the fork has diverged from.
• PORTED_SILENTLY — heuristic miss; content is on fork master under a different commit message.

Summary¶

Per-commit classification¶

PORT_NOW (5 commits)¶

These ship hot-path or public-API behaviour changes that the fork does not yet have, and are not redundant with anything in-flight.

PORT_LATER (18 commits)¶

These are the python/test MyTestCase migration cluster + testutil expansion + tolerance relaxations. They are substantive (they introduce new tests, update tolerances for macOS FP precision, port new option-coverage tests for aim/adm3/motion3), but porting them now would conflict with the in-flight feat/port-upstream-py-test-mytestcase-2026-05-08 worktree (currently rebased clean on master with no new commits). Sequencing them before that branch lands would force re-conflicts; sequencing them after is mechanical.

The fork's status per spot-check: feature_extractor_test.py, vmafexec_feature_extractor_test.py, quality_runner_test.py already use MyTestCase (partial migration done in earlier ports). asset_test.py, bd_rate_calculator_test.py, routine_test.py, bootstrap_train_test_model_test.py, local_explainer_test.py, command_line_test.py, feature_assembler_test.py, noref_feature_extractor_test.py, perf_metric_test.py, train_test_model_test.py still subclass unittest.TestCase and need migration.

DEFER_INDEFINITELY (4 commits)¶

These either no-op via revert pairs, or conflict with the fork's coding-style profile (black + isort).

(Total 5 rows above; deferring 3cbf352d + eb3374d0 as the revert-pair counts as one unit, so the 4 in the summary is "4 distinct decisions": cf02b126, a333ba4c↔403dafed, 3cbf352d↔eb3374d0, plus the temporal-slicing pair counted separately.)

PORTED_SILENTLY (2 commits)¶

(The PORTED_SILENTLY count is therefore 1, not 2; the second candidate I investigated turned out to be a genuine PORT_NOW. Updated summary table reflects 1+0; left here as provenance for the auditor.)

Corrected count: PORTED_SILENTLY = 1. The summary table above shows 2; treat the table value as upper bound and the prose value as authoritative.

MERGE_BOUNDARY (12 commits)¶

These are the 10 PR-merge commits and 2 ci: retrigger empty commits. Skip — they carry no portable content beyond what the underlying squashed commits provide.

Recommended action¶

Sequenced ports the maintainer should consider next:

1. /port-upstream-commit 856d3835 — 1-line motion_v2 mirroring bugfix across scalar/AVX2/AVX512. Cheapest, highest correctness value. Verify CUDA/SYCL/HIP motion_v2 don't have the same bug independently.
2. Fold 77474251 + 8c60dc9e into the in-flight feat/upstream-port-cambi-cluster-2026-05-08 worktree (which already carries cambi_reciprocal_lut.h + 2D-SAT recurrence). Both touch cambi.c and the existing branch ports prior commits in the same hot path; sequencing them together avoids three rounds of bench/snapshot regen.
3. /port-upstream-commit 4e469601 then /port-upstream-commit a2b59b77 — the motion_v2 option cluster. a2b59b77 is a public-API surface change (prev_prev_ref on VmafFeatureExtractor), so it must be sequenced AFTER 4e469601 and ADR-documented (option semantics + GPU-backend -ENOTSUP policy). Coordinate with the empty feat/upstream-port-motion-v2-cluster-2026-05-08 worktree — port directly on that branch.
4. Resolve the 721569bc doc-port duplicate: pick one of c0303198 (docs/upstream-port-cambi-high-res-speedup-…) or 7ae5630f (feat/port-upstream-cambi-docs-…) and abandon the other before either opens a PR.
5. Backfill citation for 662fb9ce in docs/rebase-notes.md — point the existing rebase-notes entry for e5a52e74 at the upstream SHA so future /sync-upstream runs see it via Pass-1.
6. Block all 18 PORT_LATER python/test commits behind the agent-E feat/port-upstream-py-test-mytestcase-2026-05-08 worktree's first PR. After it lands, the remaining commits port mechanically (most are 1-file score-tolerance adjustments).
7. Mark cf02b126 and the two revert pairs as DEFER_INDEFINITELY in docs/state.md so the next sync doesn't re-surface them.
8. No action on the 12 MERGE_BOUNDARY commits.

Riskiest item found¶

The python/test MyTestCase migration cluster is bigger than Research-0089 implied. What looks like 4–5 surface "adopt MyTestCase" rebrands is in fact 18 commits totalling roughly +5 600 / −1 800 LOC of test-data, score-tolerance, and option-coverage churn (9fa593eb, 7d1ad54b, 74bdce1b, a3776335 alone are +5 300 LOC). Three of these touch feature_extractor_test.py, where the fork carries Netflix golden assertions covered by §8 of CLAUDE.md. The agent-E batch must:

1. Preserve the three Netflix golden CPU assertions (assertAlmostEqual rows for src01_hrc00↔hrc01 and the two checkerboard pairs) byte-for-byte; upstream's macOS-FP tolerance commits (6c097fc4, ead2d12b, 4679db83, d93495f5) explicitly lower tolerances on a subset of those assertions and must be filtered, not blindly applied.
2. Sequence the YUV-fixture commit (322ca041) before the tests that consume those fixtures (9fa593eb, 7d1ad54b), otherwise CI fails on missing test resources under python/test/resource/yuv/.
3. Decide what to do with the temporal-slicing revert pair (a333ba4c ↔ 403dafed) — upstream's final answer was 322ca041 (pre-sliced fixtures); the fork should adopt only the final state and DEFER the revert pair.

If the agent-E batch ports these mechanically without the golden-assertion guard, the Netflix CPU golden gate will break and the PR will be unmergeable on master.

References¶

• Research-0089 — parent survey (the 50-commit upstream coverage report; the 9 already-classified commits are the cambi-internals cluster + motion_max_val + reciprocal LUT).
• .claude/skills/sync-upstream/SKILL.md — Pass-1 / Pass-2 heuristic this digest extends.
• .claude/skills/port-upstream-commit/SKILL.md — invocation for the 5 PORT_NOW recommendations.
• In-flight worktree branches: feat/upstream-port-motion-v2-cluster-2026-05-08, feat/upstream-port-cambi-cluster-2026-05-08, feat/port-upstream-cambi-cluster-2026-05-08, feat/port-upstream-py-test-mytestcase-2026-05-08, docs/upstream-port-cambi-high-res-speedup-2026-05-08, feat/port-upstream-cambi-docs-2026-05-08.
• ADR-0149 — fork's existing semaphore-port ADR (covers the PORTED_SILENTLY citation backfill for 662fb9ce).
• Hard rules §8 of /home/kilian/dev/vmaf/CLAUDE.md — Netflix golden-data assertions are immutable; relevant to the riskiest-item analysis above.