Tune xưng hô: prefer tôi/cô (tôi/anh) for same-gender rivals over coarse mày/tao

Improve subtitle merge, gender-aware xưng hô, Colab tuning, add tests

Colab: CTX_SIZE 8192 to reduce context overflow errors

Merge scene vision + translation into single call; 1 frame per cue at start

Minimal translation JSON: index+translated only, drop verbose notes

Colab: default FORCE_LLAMA_REBUILD=False to use Drive cache

Fix scene vision crash, salvage truncated JSON, retry on HTTP 400

Remove accidental full-project upload from HF repo

Clean HF repo: remove accidental full-project upload

Fix llama.cpp pin: clone tag b9553 instead of short commit ref

Fix llama.cpp pin: clone tag b9553 instead of short commit ref

Colab: MTP config (FORCE_LLAMA_REBUILD, CTX_SIZE, scene speed options)

Fix MTP: longer startup timeout, -ngld flag, pin llama.cpp b9553

Update translate_srt.py and pipeline scripts

Fix titled-name regex: honorific case-insensitive but name stays capitalized (no more 'Cô Thời ơi tỉ mỉ' corruption)

NAME GUIDE: infer character gender for titled names (fix Thầy/Cô honorific)

Update pipeline + slim Colab notebook (HF-only code source)

Add speaker-report reuse and relationship/vision Pass 2 improvements

Lock anh/em pronouns and Pass 2 multi-frame scene vision

Upload translate_srt.py with huggingface_hub

Upload translate_srt.py with huggingface_hub

Upload folder using huggingface_hub

Fix MTP detection in scripts/ensure_llama_colab.py

Fix MTP detection: use draft-mtp in --help not version number

Auto-rebuild llama-server >= b9549 for Gemma 4 MTP

Update colab/GemmaSRT_Colab.ipynb (HF_TOKEN empty)

Strip HF_TOKEN from local notebook

Update colab/GemmaSRT_Colab.ipynb

Code prep: auto-fallback drive->HF when Drive code missing

Default CODE_SOURCE=hf (always pull latest code from HF)

age/gender: robust post_init + dict tied keys; add model preflight in notebook

Fix age/gender model: post_init + dict all_tied_weights_keys

Add diarization progress logs [diarize] Step X/Y

Fix age/gender model on Colab transformers; fallback pitch, keep speaker tags

Colab: numpy/lightning fix, pyannote Agree URLs, diarize errors

Colab: pin numpy<2.1 for pyannote

Colab: verify models + pyannote before pipeline

Fix colab/GemmaSRT_Colab.ipynb JSON (HF_TOKEN sanitize)

Fix broken notebook JSON (HF_TOKEN sanitize)

Colab: simplify HF_TOKEN — paste manually in config cell

Docs: how to open Colab when repo is private (401 fix)

Colab: ENABLE_DIARIZE=True by default

Colab: HF_TOKEN auth for private repo (Secrets or config cell)

Add Pass 0 speaker diarization (age/gender) to Colab + secret-free token loading

Clarify notebook docs: SKIP_CORRECTION default True, pipeline table

Clarify SKIP_CORRECTION notes; default skip OCR

Default SKIP_CORRECTION=True (skip OCR by default)

Sync latest translate_srt, colab notebook, ensure_llama_colab

Loosen subtitle budget; disable Pass 2b shorten by default

Merge adjacent cue pairs for wider dubbing timestamps