docs: update AGENTS.md guidelines and add comprehensive UI/UX implementation plan

fix: revert to Gradio 5.0.0 due to gradio-huggingfacehub-search dependency conflict

revert: cancel Gradio v6 migration - dependency conflict with gradio-huggingfacehub-search

feat: upgrade to Gradio 6.x

fix: add missing meeting_summarizer module to Dockerfile for HF Spaces deployment

feat: initial tiny-scribe deployment

Keep only Granite-107M embedding model

Add embedding model analysis documentation

Remove Qwen2.5 3B - keep only 1.5B as extraction model

Fix Qwen2.5 3B filename pattern to lowercase

Replace Qwen2.5 Coder 3B with standard Qwen2.5 3B Instruct

Clean extraction model list: keep only Qwen2.5 models

Fix Gemma-3 1B filename pattern

Remove Granite 3.x models from extraction list

Remove models under 1B and over 2B from extraction list

Fix Gemma-3 1B repository name

Add Gemma-3 1B IT QAT to extraction models

Add E2E test results with Qwen2.5 1.5B extraction

Fix Qwen2.5 1.5B filename pattern

Add Qwen2.5 1.5B as default extraction model

Add Granite-4.0 1B model for extraction

Add E2E test script for extraction pipeline

Remove reasoning models from extraction list

Set Granite-4.0 350M as default extraction model

Add extraction benchmark and results for small models

fix: loosen schema validation and add examples to extraction prompts

feat: add Qwen2.5 3B and Qwen3 4B extraction models

fix: update Granite 4.0 Tiny 7B repo config

fix: use correct Granite 4.0 Tiny GGUF repo and set as default

fix: correct Granite 4.0 Tiny repo_id - remove full URL

fix: correct Granite 4.0 Tiny repo ID to use GGUF version

fix: correct Granite 4.0 Tiny repo ID

feat: add Granite 4.0 Tiny 7B to extraction models

docs: update analysis - Qwen3 still fails on windows >1600 tokens

fix: remove broken retry logic and lower Qwen3 extraction temperature

fix: improve extraction success rate with Qwen3 models

fix: add examples to extraction prompts to prevent empty JSON output

feat: enrich JSON export with rich debug info for all 3 pipeline stages

fix: prevent models from copying schema descriptions as extracted content

fix: make extraction robust to noisy/repetitive transcripts

fix: add response_format parameter to force JSON output from LFM2-Extract

refactor: update LFM2-Extract prompts to match official format

fix: rewrite LFM2-Extract prompt with explicit JSON format

fix: use model's actual tokenizer instead of naive character counting

fix: reserve tokens for system prompt and generation in extraction windows

debug: add print statement to capture LFM2-Extract output

debug: add detailed error logging for JSON parsing failures

fix: correct extraction count calculation in trace summary stats

fix: correct embedding parameter name in EmbeddingModel.load()

fix: update Granite embedding models to use bartowski GGUF repos