[beveradb]

Summary

Live 2026-11-05: a singer typed "Cole Phillips – Roll of Dice", a 2025 song with no karaoke version. gen's Gemini fallback "corrected" it to "Roll of the Dice", and the singer tapped keep what I typed. The make wizard then ran gen's AI judge again. The judge said "Roll of the Dice" once more, and applyFullVerdict applied it without a new audio search, because the results for the typed title were good (tier 2: Spotify "Roll of Dice", score 1.0). gen job f01c9997 ended up with the wrong title on the right audio.

Changes

• sing.js makeIsKeptTyped(): after "keep what I typed", if the form still holds what the singer typed, the wizard starts trusted. Only the formatting tidy runs, not the AI judge.
• make.js applyFullVerdict: a confident content (different-song) correction is only offered ("Did you mean…?") when the audio results already match the typed title (tier 1–2). On weak results (tier 3) it is applied with a new audio search, as before. Accepting the offer re-searches (existing acceptSuggestion).
• make.js toggleCorrection: undo/redo of a content correction searches again for the song now shown. A formatting tidy is the same song, so it keeps its results.
• Docs: CHANGELOG entry and SONG-IDENTIFICATION.md D14. Version 0.149.3. The test .py change triggers the restart that busts ?v=.

Companion PRs: karaoke-decide#207 (live Spotify search) and karaoke-gen#1256 (grounding rule). Those fix the source of the bad correction. This PR makes the wizard safe whatever the judge says.

Testing

• New E2E (tests/e2e/test_sing_make_wizard.py):
  • the incident flow: resolve correction → keep typed → wizard runs only the fast check and keeps "Roll of dice";
  • strong tier + content verdict → offered, title unchanged, no second search, and accepting re-searches;
  • weak tier → applied and re-searched; undo → back to typed with its own search and no extra judge call.
• With each fix reverted, its new test fails (checked).
• Make-wizard E2E file: 41 passed. Full suite on the build host (scripts/kjbox-test-remote): all pass except test_routes.py::test_download_queues_item / test_download_failure_tracked, which fail the same way on main there (host environment).
• Local CodeRabbit review: no findings.

@coderabbitai ignore

🤖 Generated with Claude Code

https://claude.ai/code/session_012RYLVUQ68KFYtZF7vEfhAq