Moviola Protocol MCP 0.3.4 - Progress Narration and Closed-Stdin Fix

Reliability patch for long conversions over MCP. Two intertwined fixes, both found by measurement in live use: (1) ffmpeg/ffprobe children inherited the MCP stdio pipe as stdin, so a 1.7-second encode could hang an MCP call past 20 seconds while the conversion itself completed and indexed correctly - every core subprocess now pins stdin=DEVNULL, with unit tests guarding the invariant; (2) the long tools (convert_media, slideshow_from_images, export_training_data) now narrate themselves via FastMCP Context: progress 0.0 at start, staged heartbeats while ffmpeg works, 1.0 when the index row lands, plus info logs. README documents Progress & timeouts: on any client timeout, probe the output and read index_history before re-running - the ledger row is the completion certificate. Verified: 45 tests green, verify_all five gates pass, real stdio progress session measured (events 0.0 -> 1.0, 2.3 s end-to-end where the pre-fix path timed out at 20 s).

Moviola Protocol MCP 0.3.3 - Complete AGPL and EUPL Licence Texts

Licence-completeness patch. The root short notice is now LICENSE.md, and the complete canonical GNU Affero General Public License v3 and European Union Public Licence 1.2 texts ship byte-identically as LICENSE-AGPL-3.0.txt, LICENSE-EUPL-1.2.txt, LICENSES/AGPL-3.0-or-later.txt, and LICENSES/EUPL-1.2.txt, mirroring RTR. AGPL is 34,523 bytes (SHA-256 0d96a4ff68ad6d4b6f1f30f713b18d5184912ba8dd389f86aa7710db079abcb0); EUPL is 13,792 bytes (SHA-256 57fb42fbcd0b037ce528ed8f72f1ec095d67bc6825ecf1448ff39be1fe68a4b4). Local Licensee 10.1.0 detects AGPL-3.0 by exact 100% match and EUPL-1.2 at 99.1957% Dice confidence. Python metadata moves to PEP 639 License-Expression AGPL-3.0-or-later OR EUPL-1.2 with all five files declared and present in the wheel and sdist. scripts/verify_all.py now audits exact licence sizes/hashes and Licensee detection. No functional conversion changes. 41 tests and all five verification gates pass.

Moviola Protocol MCP 0.3.2 - Post-Scriptum for Lecturers

Documentation release: docs/post-scriptum/Post-Scriptum.md, a lecturer-facing essay and teaching kit (what the release means for AI training, measured vs reasoned vs unknown, the MMPROJ question as an affordable experiment, a four-session module, five labs with measured expected outputs, discussion questions), plus its dedicated Post-Scriptum zone in the README served identically on GitHub and PyPI. No code changes from 0.3.1. Quintuple check: all five gates pass.

Moviola Protocol MCP 0.3.1 - Expanded Presets, HOW TO TRAIN, Witnesses, Verbose

Five full-width 30-second preset representations (Passthrough baseline 578,695 B). HOW TO TRAIN: Special Datasets Philosophy, 20-camp atlas, worked recipes (visemes, documentaries, bio/geo, tomes 331-1690 B/page). Hard witness bench in docs/witnesses (57 rows, integrity clean, byte-identical repro). Permanent verbose tracing and scripts/verify_all.py quintuple check - gate 3 caught and fixed unclosed SQLite connections. Still-image GIF bug fixed. 41 tests + 5 gates green.

Moviola Protocol MCP 0.3.0 - Scaling Ladder, TorchPy Backends, Premier

Resolution becomes a named choice: --dpi (pixels per inch at a stated reference width) joins --height across core, CLI and MCP (11 tools). Three scaling-ladder GIFs (DPI / p / Torch variants, stages 1080-720-480-350-240-120) with indexed stage tables; the 720p 1-bit anomaly is published, not smoothed. TorchPy bridge: ladder-native MoviolaFrameDataset for all five renderings, Tessersect unpack, image-folder/WebDataset/ArrayRecord/Unsloth exports, fps-passthrough frame integrity (a 334-from-167 duplication bug, measured and regression-tested). Training backends ring for non-NVIDIA hosts: PyTorch ROCm, AITER accelerated attention + a8w8 GEMM dispatch with stock fallback, JAX arrays, MaxText ArrayRecord, TorchTitan and Megatron-LM interop formats, AMD Primus datacenter shards + manifest, Unsloth vision-SFT extra - moviola backends reports this host honestly. Original papers moved to sources/ with catalog; README gains the Premier (Symbiose, announced not shipped). 38 tests green. PyPI upload still pending a project token. El Psy Kongroo.

Moviola Protocol MCP 0.2.0 - Canine Logic Moviola Tessersect

Preset D arrives: Canine Logic Moviola Tessersect - 90 fps sampling in the canine flicker band, 120p one-bit Grey-Ink cells, sixteen instants collaged into one 4x4 tesseract-net frame for high-velocity absorption during training (65,525 B on the 29.6 s reference clip). README rewritten in Nova-Violet-Role org style, fully technical for PyPI readers, with 30-second demo GIFs per preset. Easter egg: moviola flash (Intersect flash + worldline divergence computed from moviola.db; docs/easter-egg.md). Copyright Saimonokuma, AGPL-3.0-or-later OR EUPL-1.2. 21 tests green; wheel verified by content listing and clean-room uvx run. Install: uvx --from moviola-protocol-mcp. PyPI upload still pending a project token (name is free). El Psy Kongroo.

Moviola Protocol MCP 0.1.0

First release. MCP server (stdio) + CLI converting any video/image format through three measured Moviola grey presets (Standard NTSC-J 8-bit, Intersect 1-bit overcalc, Torch 5fps 360p/480p), every conversion indexed in moviola.db. Bundles the gif-as-mp4 OpenCode plugin (.mjs + .ts). Install: uvx --from moviola-protocol-mcp, or uvx --from git+https://github.com/Nova-Violet-Role/Moviola-Protocol-MCP moviola-protocol-mcp. PyPI upload pending token (name moviola-protocol-mcp is free).