Fix LLM2Vec PEFT adapter stacking

ZeroGPU-safe startup device + llm2vec tensor warning cleanup

fix: prefer local text encoder API in api mode when available

fix: fallback to local text encoder API when remote URL times out

fix: HTTP readiness gate in _ensure_text_encoder_server

fix: lazy TextEncoderAPI client with retry + HTTP readiness gate

Fix text encoder payload parsing and add remote encoder switch

Switch to LLM2Vec 3.1 pair to fix Space 401 on gated Llama 3.0

Sync text encoder 3.0 fix and smoke checks

Fix HF auth propagation for text encoder loads

Switch LLM2Vec to Meta-Llama-3.1-8B-Instruct (accessible with current token)

Improve error extraction: collect all error strings from text encoder response

Use mirrored LLM2Vec base model repo and env overrides

Debug: print full traceback on text encoder init failure to expose real error

Improve text encoder error messages: better diagnostics for HF token issues

Improve text encoder API error detection: validate .npy paths and detect error messages

Fix generate notification: reuse persistent notification and properly transition to error state

Use lazy TextEncoderAPI client in API-only mode (no startup fallback)

Add core kimodo package modules required by native demo