docs(ai): record measured retrain gate status, not authoring-time status

[lusoris]

Summary

The epic #1246 gate table listed G2 and G3 as FAIL against conditions that no longer hold — G3 in particular blamed "PR #1307 & fix/cambi-cuda-context unmerged", and both merged on 2026-09-05/06. That is the table the retrain is supposed to be gated on, so a stale FAIL is as costly as a wrong PASS.

Each gate was re-run, not reasoned about:

Gate	Was	Now	Why
G1	FAIL	FAIL	12 epics open, now listed by number, with the caveat that the bodies are snapshots and several items have already shipped
G2	FAIL	PASS	Zero failing runs on master HEAD. The release-please failure it cited is gone — ADR-1171 made the missing release-bot credential warn-not-error on push
G3	FAIL	PASS	#1307, #1312, #1324 all on master; container rebuilt from cd52f2670, default model verified on all four backends
G4	FAIL	FAIL	Still #1302 — now with the precise reason
G5	FAIL	FAIL	Maintainer sign-off, unchanged

G3 was checked more broadly than the runbook asked. §3.2 specifies CUDA only. A model reaches the GPU twins through the model, not through --feature, so a CUDA-only check would not have covered SYCL or HIP at all. All four were run: CPU 82.816062, CUDA 82.814062, SYCL 82.814061, HIP 82.816061, each exiting 0 — evidence and container digest.

G4's blocker is now specific. master's extract_k150k_features.py has no --vmaf-model flag (grep -c returns 0), which is exactly what §4.2's teacher_model assertion needs. #1302 supplies it, and #1302 has no failing check — its only red mark is the aggregator's draft guard (Draft PRs must not satisfy Required Checks Aggregator), and its ADR-0108 validator passes six of six. It needs promotion, not repair.

Type

• docs — documentation only

Checklist

• Commits follow Conventional Commits.
• make format && make lint green locally — pre-commit clean on the touched files.
• Unit tests pass — no code touched; this is the retrain runbook's gate table.
• SIMD/GPU code path — none touched. The GPU numbers quoted were measured, not changed.
• Feature extractor twins — none touched.
• New C/C++ source — none added.
• Breaking change — no.
• ADR — none: this records measurements. No decision is made that another engineer could have made differently.

Bug-status hygiene (ADR-0165)

• docs/state.md updated — no state delta: gate status is epic bookkeeping, not a tracked defect.

Netflix golden-data gate (ADR-0024)

• I did not modify any assertAlmostEqual(...) score.

Deep-dive deliverables (ADR-0108)

• Research digest — no digest needed: trivial. Five verification commands re-run and their results recorded.
• Decision matrix — no alternatives: only-one-way fix. A gate is either passing or not; the table either says so accurately or it does not.
• AGENTS.md invariant note — no rebase-sensitive invariants: the one thing a future editor must not do (carry a status forward without re-running it) is stated in docs/rebase-notes.md and inline in the table's own note.
• Reproducer / smoke-test command — below.
• CHANGELOG fragment — changelog.d/changed/retrain-gate-status-1246.md.
• Rebase note — docs/retrain-gate-status-1246 — measured retrain gate status (2026-09-06).

Reproducer

# G2 — zero failing runs on master HEAD
gh run list --branch master --limit 20 --json conclusion,name \
  --jq '[.[]|select(.conclusion=="failure")]'

# G4 — the flag §4.2 needs is absent from master
git show origin/master:ai/scripts/extract_k150k_features.py | grep -c 'vmaf.model'   # 0

# ...and #1302 supplies it, with no failing check of its own
gh pr checks 1302 --json name,state --jq '[.[]|select(.state=="FAILURE")]'

Known follow-ups

• G1 needs an audit, not a count. Twelve open epics overstate the work: the bodies are snapshots and several items have already shipped. Each should be checked against the code before being treated as outstanding — the same drift this PR just corrected in G3.
• G4 unblocks the moment feat(ai): distil from the ADR-1168 default model with per-row teacher provenance #1302 is promoted, after which §4 can run and the gate can be re-measured.

[lusoris]

Extended: this PR now also fixes a path that would have aborted the retrain.

The runbook pointed --scores at .corpus/konvid-150k/scores.csv, which does not exist — KoNViD-150k splits its scores into k150ka_scores.csv / k150kb_scores.csv. The wrong path appeared twice: in the §4 five-clip smoke and in the §5.1 multi-day production extraction. extract_k150k_features.py validates it and exits error: scores CSV not found, so the smoke would have died on its first line and the ~105–110 h K150K run at its very start.

Verified against the workstation corpus:

k150ka_scores.csv   154,746 rows   <- exists, and the script's own default
scores.csv          absent
clips/              153,841 files  <- superset of k150ka_extracted/ (152,265)

Corrected to k150ka_scores.csv. --clips-dir is left as clips/ because lookup is by video_name and either directory resolves; both paths are now in a verify-first command so this is caught before the long run rather than hours into it.