- Add --model-list-batch to EXAMPLES.md for batch model listing
- Add new CLI options to MCP server (--output-dir, --yes, --audio-chunk)
- Add new CLI options to webapp build_command function
- Update README.md with Output Options section

Added --audio-chunk argument with 3 modes:
- overlap (default): overlapping chunks like [0-60], [58-118]
- word-boundary: uses Whisper timestamps to split at word boundaries
- vad: uses Voice Activity Detection to skip silence

Also added --audio-chunk-overlap to control overlap duration.

New functions added:
- process_video_with_vad(): VAD-based chunking
- process_video_word_boundary(): Word-boundary chunking using Whisper

Modified:
- transcribe_video_audio(): accepts audio_chunk_type and audio_chunk_overlap params
- _transcribe_chunked(): accepts chunk_type and overlap params

Models like Faber8/AbyssOrangeMix2 are SD1.5 models, not Flux.
Now detected as StableDiffusionPipeline instead of FluxPipeline.

When a model has a VAE configured but the VAE files don't exist in
the repository, try loading with the default VAE instead.

Allow using full HuggingFace model ID (e.g., Faber8/AbyssOrangeMix2_nsfw)
as --model argument by looking up both short name and full ID in MODELS.

The repos attribute from scan_cache_dir() is a frozenset, not a list,
so .pop() doesn't work. Fixed by using next(iter()) instead.

The message is now suppressed when using --model-list-batch to make
the output cleaner for scripts.

Added --yes / -y argument that automatically answers 'yes' to confirmation
prompts when deleting cached models or clearing the entire cache.

Usage:
  videogen --remove-cached-model MODEL_ID --yes
  videogen --clear-cache --yes

The --model-list-batch option was added but wasn't being handled properly.
Now it correctly exits after printing the batch output, and is also
added to the list of options that don't require --prompt.

- Fixed model ID consistency: numeric IDs now remain the same when using
  filters like --nsfw-friendly, --t2v-only, --i2v-only, etc.
  Previously, filtered lists would renumber models making --show-model
  by numeric ID unreliable.

- Added --model-list-batch option for script-friendly output:
  Outputs 'NUMERIC_ID:FULL_MODEL_NAME' format for easy parsing

- Added --output-dir option to specify output directory:
  Sets the directory where output files will be saved

- Fixed syntax error in argparse epilog string that was causing
  'SyntaxError: invalid decimal literal'

- Fixed model ID normalization to handle hyphens (in addition to underscores)
- Fixed dictionary key ordering in base_model_fallbacks so more specific keys (wan2.2.i2v) are checked before generic keys (wan2.2)
- Fixed Wan 2.1 I2V base model mapping (was incorrectly pointing to T2V)
- Fixed base model detection in earlier code sections to check model ID directly instead of relying on m_info.get('supports_i2v')
- Fixed typo: Wan 2.2 generic fallback now correctly uses Wan2.2-T2V

Now Wan 2.2 I2V models like Wan-AI/Wan2.2-I2V-A14B will correctly use Wan-AI/Wan2.2-I2V-14B-Diffusers as the base model instead of the incorrect Wan-AI/Wan2.2-T2V-14B-Diffusers.

- Fixed mapping table to use correct I2V base model (Wan-AI/Wan2.2-I2V-14B-Diffusers)
- Fixed Diffuser -> Diffusers typo in model IDs
- Updated all Wan 2.2 I2V references to use correct model ID

- Add correct base model for Wan 2.2 I2V: Wan-AI/Wan2.2-I2V-A14B-Diffuser
- Add specific VRAM estimate for Wan 2.2 I2V MoE models (~14GB)
- Apply more conservative VRAM calculation for models with weights/LoRAs
- Fix indentation error in add_model_from_hf function

Calculate: base_vram + 2GB + 50%
This ensures a 14B model estimated at 18GB will require ~29GB instead of 22.5GB.

Instead of adding a fixed 2GB overhead, now calculates 25% more
VRAM for base models that will have fine-tuned weights/tensors
or LoRA adapters loaded on top.

The user confirmed that Wan2.2-I2V models should use Wan2.2-T2V
as the base model, not the I2V variant.

The issue was that model IDs from HuggingFace use dots (wan2.2-i2v-a14b)
while user config names use underscores (wan2_2_i2v_a14b).

Now we normalize the model ID by replacing underscores with dots before
matching against the base_model_fallbacks dictionary.

The issue was that model IDs like 'wan2_2_i2v_a14b' would match
'wan2_2' (T2V) before 'wan2_2_i2v' (I2V) because 'wan2_2' comes
first in the dictionary.

Now the dictionary is ordered with more specific keys first:
- wan2_2_i2v_a14b (most specific)
- wan2.2_i2v_a14b
- wan2_2_i2v
- wan2.2_i2v
- wan2_2
- wan2.2
etc.

This ensures longer/more specific keys are checked before shorter ones.

Before this fix, using Wan 2.2 I2V models like 'wan2_2_i2v_a14b' would
incorrectly try to load 'Wan-AI/Wan2.1-T2V-14B-Diffusers' as the base
model because the fallback logic only matched 'wan' generically.

Now the base_model_fallbacks dictionary includes specific entries for:
- Wan 2.2 I2V models: wan2_2_i2v, wan2.2_i2v
- Wan 2.2 T2V models: wan2_2, wan2.2
- Wan 2.1 I2V models: wan2_1_i2v, wan2.1_i2v
- Wan 2.1 T2V models: wan2_1, wan2.1
- Generic Wan fallback: wan

The more specific keys are checked first, so model IDs like
'wan2_2_i2v_a14b' will correctly match 'wan2_2_i2v' and use
'Wan-AI/Wan2.2-I2V-14B-Diffusers' as the base model.

- Add get_pipeline_for_task() function that determines pipeline
  based on model ID AND task type (t2v, i2v, t2i, i2i, v2v)
- Pipeline class is now ALWAYS detected at runtime, not from config
- Remove old dynamic switching code that's now redundant
- Update check_model.py to show runtime detection instead of fixing config
- Update check_pipelines.py to show V2V pipelines

- Add WanImageToVideoPipeline and WanVideoToVideoPipeline to PIPELINE_CLASS_MAP
- Add CogVideoXImageToVideoPipeline and CogVideoXVideoToVideoPipeline
- Add AnimateDiffVideoToVideoPipeline
- Add StableDiffusionImg2ImgPipeline for SD 1.5
- Add dynamic pipeline switching logic for Wan, LTX, CogVideoX, AnimateDiff
- The pipeline class is now selected at runtime based on task mode
- Fix detect_pipeline_class to correctly identify Wan models
- Remove duplicate LTX handling code

Adjust VRAM checking to allow models up to 10% less than available VRAM, or full VRAM with offload strategy