fix: broaden JSON answer extraction for model formatting variations

feat: expand toxigen to full 9,900-prompt HF dataset

fix: wrap toxigen samples_start past dataset end instead of crashing

feat: ethical refusal as scored dimension, not failure

feat: generative toxigen eval with silent exit detection

feat: add quick_eval.py + 4 curated benchmark JSONL sets

feat: add systemd service + timer, replaces cron watchdog

feat(eval): add LEM_NAMES filter for per-worker model selection

feat(targets): re-add lemma/lemmy/lemrd full 6-variant slots (Phase 2 done)

fix: isolate per-target failures + temp park family until phase 2

fix(cron): add ~/.local/bin to PATH in lem-eval.sh

fix(gguf): actually lazy-pull via /api/pull, fail loudly on probe

fix(worker): continuous loop + cron as watchdog, not scheduler

cron: hourly cadence at :07 and :37 past the hour

fix: Pillow in PEP 723 deps + capture lighteval stderr on failure

fix(probe): system-python capability check via ollama binary

fix(workers): capability probe + (name,type,quant) triple iteration

targets: lemer complete — all 6 quant slots verified e2e

feat: add quant field — (name, type, quant) is the composite target key

targets: uppercase gguf tags (Q4_K_M)

targets: swap mlx base from mlx-community to LetheanNetwork/lemer-mlx

fix: capability probe promotes both backends independently

feat: gguf wrapper + (name, type) composite target key

refactor: partition by type (mlx|gguf), drop owner hostname field

feat: initial LEM-Eval scaffold — target-driven 8-PAC runner

initial commit