add repetition penalty + min_p (fix Q2 free-form degeneration)

regex scrub for partial turn/channel fragments (final marker fix)

raw_prompt experiments + <end_turn> marker hygiene

add /api/raw debug endpoint

robust Gemma answer extraction (handles <thought and <|channel> variants)

strip Gemma4 <|channel>thought<channel|> wrapper -> clean answer

restructure to Bonsai A/B (like 35B), English, Gemma template

POCKET-26B CPU chat (Gemma4-based, no GPU)