chore(api): remove /classify/debug diagnostic endpoint

feat(api): add /classify/debug endpoint for stage-by-stage diagnostics

revert: go back to in-Space inference (free CPU on Space)

fix: switch to router.huggingface.co/hf-inference (api-inference.* deprecated)

refactor: thin HF Inference API proxy for v5 models

chore: suppress transformers warnings via env var, clean up app.py

chore: suppress beta/gamma and sklearn version warnings

fix: load stage3 tokenizer from stage1a subfolder (stage3/tokenizer.json is incompatible)

fix: use use_fast=False for stage3 tokenizer to bypass tokenizer.json

fix: use AutoTokenizer for stage3 to fix tokenizer.json version mismatch

feat: add Longformer pipeline with lazy-load model registry

fix: build model offline from BertConfig + .pt weights

use mental-bert with HF token auth

initial deployment