ltx-fidelity-debugging

# LTX Fidelity Debugging Use this skill for the narrow slice where the real `LTX-2.3` path runs but the output is still visually wrong, ambiguous, or not strong enough to promote into repo truth. This skill is for: - real-weight local smoke runs - stage-local visual debugging - safe iteration on the `22B` AV path - deciding whether the next suspect is prompt/stage 1, x2/stage 2, or decode/export Do not use this skill for the common implementation path. The default repo loop still lives in `AGENTS.md`, `MEMORY.md`, and `scripts/dev.py`. ## Default workflow 1. Read `docs/research/10-ltx-fidelity-debug-playbook.md`. 2. Use `scripts/ltx_debug_smoke.py` with a fixed seed and real converted artifacts. 3. Start with the named profile ladder: `safe-smoke`, then `visual-gate`, then `coherence`. 4. Keep clean lifecycle on unless you are disproving the lifecycle itself. 5. Use detached `tmux` execution for long real-weight smokes. 6. Inspect `stage1`, `post_x2`, and `final` snapshots before deciding what to fix next. 7. Review `frame_0001`, `frame_mid`, and `frame_last` before promoting a rung. 8. Downgrade the claim if the result is ambiguous. ## What to load - `docs/research/10-ltx-fidelity-debug-playbook.md` - `docs/research/07-ltx-integration-seams.md` - `docs/research/09-open-questions-and-validation-plan.md` - `MEMORY.md` Load these code paths only when you need the implementation seam: - `packages/families/ltx/src/ltx/_prompt_encoding_backend.py` - `packages/families/ltx/src/ltx/_generation_backend.py` - `scripts/ltx_debug_smoke.py` ## Rules - Start at the smallest truthful rung, not the highest resolution. - Use real converted artifacts for visual conclusions. - Change one major variable at a time. - Trust stage-local evidence over the final clip when localizing a bug. - Generic quadruped output does not count as dog success.