[kaffeindecaf]

changes

• add causality inversion with bounded future-frame blend, replacement, band tearing and motion alignment
• expose lookahead, offset, strength, region masks, speed, alignment, feather and seed in CLI/menu
• preserve original rational video timestamps through the shared renderer with optional PyAV/NUT transport
• copy compatible unfiltered audio, drain at EOF/cuts, and keep the queue under 128 frames/128 MiB
• document latency, live-input limits, dense-VFR/EOF fallback, optional dependencies and transient memory
• check off E2.49 and save separate mode/resolution benchmarks
• add a GitHub Actions regression gate with FFmpeg and all optional processors installed

verification

• all ten local test suites pass
• 12 focused inversion tests, including real VFR and nonzero-start timestamp equality
• copied AAC packet timestamps match the source
• early moving-object appearance, distinct modes, scene cuts, EOF fallback, deterministic seeds and bounded state
• sediment/collapse in both chain orders; FFmpeg filters before and after
• short pipe writes, encoder failure and Ctrl-C remove partial output
• menu control/render smoke test
• eight 360p/720p benchmark cases in benchmarks/inversion.json
• Ruff F gate, added-line security scan and independent code review

GitHub CI must pass before merge. Benchmarks measure processor work, not end-to-end decode/encode throughput. Stateful previews and separated stateful groups remain unsupported.