feat: improve knowledge extraction and wire knowledge labels

fix: restore knowledge extraction and token tracking

feat: multi-provider LLM (Ollama local + HuggingFace deployed + Groq), fix think-tag parsing, 15K batch

fix: 5K char batch — Groq free tier strict payload limit

restore: gpt-oss-120b model, 30K batch size, proper system prompt

debug: reduce max_tokens to 2048, add request size logging

fix: minimal prompt + 3K batch for Groq limits, optimistic delete in UI

fix: switch to gpt-oss-20b (faster, larger limits), reduce batch to 8K

fix: reduce batch size to 20K chars for Groq free tier TPM limits

fix: resolve LLM context limits and tool-calling errors using Qwen JSON batching

feat: batch knowledge extraction — processes all sections sequentially, no data loss

refactor: knowledge extraction uses native JSON mode, no tool calling, 50K char buffer

fix: reduce prompt size aggressively (6K cap, 4K retry, 3K fallback) for Groq limits

fix: cap knowledge extraction prompt to 12K chars, reduce retry to 3 sections

feat: retry failed docs, graceful shutdown, optimistic approve, timeline fix, dashboard latest workflow

feat: auto-resume orphaned workflows on server restart + cancel button

feat: live stage progression during workflow execution

fix: capture token counts from LLM using include_raw=True

feat: end-to-end cost tracking with frontend metrics panel

feat: upgrade to GPT OSS 120B + rewrite extraction prompt

feat: add LLM-powered knowledge extraction workflow