feat(ci): nightly bisect-model-quality + sticky tracker (closes #4)

[lusoris]

Summary

• Wires the existing vmaf-train bisect-model-quality tool into a nightly GitHub Actions workflow (.github/workflows/nightly-bisect.yml) that runs against a deterministic synthetic placeholder cache (committed under ai/testdata/bisect/, regenerable via ai/scripts/build_bisect_cache.py).
• Always edits a single sticky comment on tracker issue #40 with the verdict + per-step PLCC/SROCC/RMSE table; fails CI red on first_bad_index is not None.
• Closes Wire bisect-model-quality into a nightly CI workflow #4. Real DMOS-aligned cache + canonical model timeline ordering swap in via a follow-up — see ADR-0109 + Research-0001 for the swap path.

Type

• feat — new feature
• ci — tooling / infra

Checklist

• Commits follow Conventional Commits (feat(ci): …).
• No SIMD/GPU code touched — /cross-backend-diff not applicable.
• No feature extractor SIMD/GPU twins touched.
• New .py files carry the Copyright 2026 Lusoris and Claude (Anthropic) header.
• Not a breaking change.

Netflix golden-data gate (ADR-0024)

• No assertAlmostEqual(...) scores in Netflix golden Python tests modified.

Cross-backend numerical results

Not applicable — no SIMD/GPU code touched.

Performance

Not applicable — net cost is one nightly job (Python-only, ~15 min budget).

Deep-dive deliverables (ADR-0108)

• Research digest — docs/research/0001-bisect-model-quality-cache.md.
• Decision matrix — captured in ADR-0109 ## Alternatives considered (6 options, 5 alternatives).
• AGENTS.md invariant note — added byte-stable-cache invariant to ai/AGENTS.md.
• Reproducer / smoke-test command — see below.
• CHANGELOG.md "lusoris fork" entry — bullet under ### Added (Tiny AI nightly bisect).
• Rebase note — entry 0011 — Nightly bisect-model-quality + fixture cache in docs/rebase-notes.md.

Reproducer

# Local smoke (matches what the workflow does):
python ai/scripts/build_bisect_cache.py --check
vmaf-train bisect-model-quality \
    ai/testdata/bisect/models/model_*.onnx \
    --features ai/testdata/bisect/features.parquet \
    --min-plcc 0.85 --input-name input \
    --json /tmp/bisect-result.json --fail-on-first-bad
# Expected: "no regression in this range"; first_bad_index = None; exit 0.

# Verify the synthetic-regression case (existing pytest):
pytest ai/tests/test_bisect_model_quality.py::test_bisect_localises_first_bad -v

Known follow-ups

• Replace the synthetic placeholder cache with a real DMOS-aligned subset (NFLX-public is the likely first target). The workflow file + sticky-comment helper stay unchanged at swap time.
• Optionally retire the synthetic cache once a real one lands, or keep both as separate jobs (synthetic = wiring smoke, real = quality regression detection). Lean toward keeping both.

🤖 Generated with Claude Code