2 months ago
c3eebe1
Davy* chore: rebase bounded separation onto current main * chore: apply final bounded separation correction * fix(separator): require named two-stem outputs * chore: retrigger CI after separator correction * chore: restore runtime fixture in streaming tests * test(separator): initialize runtime in streaming inference tests * chore: retrigger CI after runtime fixture correction * fix(separator): keep short audio on one inference window * fix(separator): align reported chunk totals with the streaming loop The long-input chunk total undercounted the streaming loop iterations (step_by covers every hop start, not just the coverage minimum), so on_chunk_complete could exceed the reported total and push progress past the stage bound. chunk_schedule now derives the total directly from the loop schedule. Also extracts classify_output_contract from detect_output_contract and adds the regression tests requested in review: generic output_0/output_1 two-output models are rejected, named vocals/accompaniment bundles are accepted, and short audio stays on one inference window. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> --------- Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Parent030053c