Wire bisect-model-quality into a nightly CI workflow

[lusoris]

Goal

The bisect-model-quality tool landed in commit 4a6b76eb (feat(ai): bisect-model-quality — binary-search a checkpoint timeline), but it's
only usable locally today. To make it load-bearing we should run it
nightly against our model registry and alert on regressions.

What's needed

1. Golden feature cache — a committed or cached parquet of feature
   vectors + DMOS targets drawn from a frozen subset of NFLX-public /
   LIVE / KonIQ. Must be stable across CI runs; probably lives in
   testdata/ or ai/testdata/.
2. Nightly workflow (.github/workflows/nightly-bisect.yml):
   • Checkout, install ai/ package + onnxruntime
   • Run vmaf-train bisect-model-quality --models model/*.onnx --features <cache> --min-plcc 0.85
   • On first_bad_index is not None, post a comment to a tracking
     issue (not file a new one each time — use a sticky label + edit-
     in-place)
3. Registry ordering — the tool assumes monotonic quality on the
   model list. We need a canonical ordering (git log on model/*.onnx?
   release tags?) that maps models → timeline indices reproducibly.

Why this is deferred

• Requires designing the golden feature cache first (item 1 above is
  non-trivial: needs to be reproducible, not too big for git, aligned
  to the FR/NR regressor input shape).
• We don't have a stable model registry cadence yet — tiny-AI model
  releases are still in flux.
• Other CI items (see .workingdir2/analysis/ci-security-triage.md)
  are higher prio (P0/P1 vs this being P2).

Acceptance criteria

• Nightly workflow runs and passes on a known-good model set
  (first_bad_index is None).
• Synthetic regression test: introducing a deliberately bad ONNX into
  the set trips the alert.
• Report is readable (render_table() output is posted to the tracking
  issue).
• No false positives from stochastic eval (fixed RNG seeds in feature
  cache).

Out of scope

• Training the cache-generation script itself (separate tiny-AI task).
• Per-commit bisection (would 10x CI cost — nightly is the right cadence).